diff --git a/custom_nodes/ComfyUI-MMAudio/.gitattributes b/custom_nodes/ComfyUI-MMAudio/.gitattributes new file mode 100644 index 0000000000000000000000000000000000000000..dfe0770424b2a19faf507a501ebfc23be8f54e7b --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/.gitattributes @@ -0,0 +1,2 @@ +# Auto detect text files and perform LF normalization +* text=auto diff --git a/custom_nodes/ComfyUI-MMAudio/.github/workflows/publish.yml b/custom_nodes/ComfyUI-MMAudio/.github/workflows/publish.yml new file mode 100644 index 0000000000000000000000000000000000000000..1ad82c7afd489bd1e3bb2c8e2c492d33e3f462b2 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/.github/workflows/publish.yml @@ -0,0 +1,26 @@ +name: Publish to Comfy registry +on: + workflow_dispatch: + push: + branches: + - main + - master + paths: + - "pyproject.toml" + +permissions: + issues: write + +jobs: + publish-node: + name: Publish Custom Node to registry + runs-on: ubuntu-latest + if: ${{ github.repository_owner == 'kijai' }} + steps: + - name: Check out code + uses: actions/checkout@v4 + - name: Publish Custom Node + uses: Comfy-Org/publish-node-action@v1 + with: + ## Add your own personal access token to your Github Repository secrets and reference it here. + personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }} diff --git a/custom_nodes/ComfyUI-MMAudio/.gitignore b/custom_nodes/ComfyUI-MMAudio/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..d75870d1b348a0603e9dd937344d2a1ee3797575 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/.gitignore @@ -0,0 +1,11 @@ +output/ +*__pycache__/ +samples*/ +runs/ +checkpoints/ +master_ip +logs/ +*.DS_Store +.idea +*.pt +tools/ \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/LICENSE b/custom_nodes/ComfyUI-MMAudio/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..086f821dd945c2f8ab0fc804063139253ccfa22b --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2025 Jukka Seppänen + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/custom_nodes/ComfyUI-MMAudio/__init__.py b/custom_nodes/ComfyUI-MMAudio/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f4e09a981f5d04fa9bca70cd103131775e9514da --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/__init__.py @@ -0,0 +1,4 @@ +from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS + + +__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"] \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/configs/DFN5B-CLIP-ViT-H-14-384.json b/custom_nodes/ComfyUI-MMAudio/configs/DFN5B-CLIP-ViT-H-14-384.json new file mode 100644 index 0000000000000000000000000000000000000000..a1b1df6bd4addb1ce6be5b8af094c3dfc5e1bee7 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/configs/DFN5B-CLIP-ViT-H-14-384.json @@ -0,0 +1,34 @@ +{ + "model_cfg": { + "embed_dim": 1024, + "quick_gelu": true, + "vision_cfg": { + "image_size": 378, + "layers": 32, + "width": 1280, + "head_width": 80, + "patch_size": 14 + }, + "text_cfg": { + "context_length": 77, + "vocab_size": 49408, + "width": 1024, + "heads": 16, + "layers": 24 + } + }, + "preprocess_cfg": { + "mean": [ + 0.48145466, + 0.4578275, + 0.40821073 + ], + "std": [ + 0.26862954, + 0.26130258, + 0.27577711 + ], + "interpolation": "bicubic", + "resize_mode": "squash" + } +} \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/configs/config.json b/custom_nodes/ComfyUI-MMAudio/configs/config.json new file mode 100644 index 0000000000000000000000000000000000000000..cdb2136fb5c684c0539e8ef0e65fb89a3d3af5e6 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/configs/config.json @@ -0,0 +1,165 @@ +{ + "_commit_hash": null, + "architectures": [ + "CLIPModel" + ], + "initializer_factor": 1.0, + "logit_scale_init_value": 2.6592, + "model_type": "clip", + "projection_dim": 1024, + "text_config": { + "_name_or_path": "", + "add_cross_attention": false, + "architectures": null, + "attention_dropout": 0.0, + "bad_words_ids": null, + "begin_suppress_tokens": null, + "bos_token_id": 0, + "chunk_size_feed_forward": 0, + "cross_attention_hidden_size": null, + "decoder_start_token_id": null, + "diversity_penalty": 0.0, + "do_sample": false, + "early_stopping": false, + "encoder_no_repeat_ngram_size": 0, + "eos_token_id": 49407, + "exponential_decay_length_penalty": null, + "finetuning_task": null, + "forced_bos_token_id": null, + "forced_eos_token_id": null, + "hidden_act": "quick_gelu", + "hidden_size": 1024, + "id2label": { + "0": "LABEL_0", + "1": "LABEL_1" + }, + "initializer_factor": 1.0, + "initializer_range": 0.02, + "intermediate_size": 4096, + "is_decoder": false, + "is_encoder_decoder": false, + "label2id": { + "LABEL_0": 0, + "LABEL_1": 1 + }, + "layer_norm_eps": 1e-05, + "length_penalty": 1.0, + "max_length": 20, + "max_position_embeddings": 77, + "min_length": 0, + "model_type": "clip_text_model", + "no_repeat_ngram_size": 0, + "num_attention_heads": 16, + "num_beam_groups": 1, + "num_beams": 1, + "num_hidden_layers": 24, + "num_return_sequences": 1, + "output_attentions": false, + "output_hidden_states": false, + "output_scores": false, + "pad_token_id": 49408, + "prefix": null, + "problem_type": null, + "projection_dim": 512, + "pruned_heads": {}, + "remove_invalid_values": false, + "repetition_penalty": 1.0, + "return_dict": true, + "return_dict_in_generate": false, + "sep_token_id": null, + "suppress_tokens": null, + "task_specific_params": null, + "temperature": 1.0, + "tf_legacy_loss": false, + "tie_encoder_decoder": false, + "tie_word_embeddings": true, + "tokenizer_class": null, + "top_k": 50, + "top_p": 1.0, + "torch_dtype": null, + "torchscript": false, + "transformers_version": "4.27.1", + "typical_p": 1.0, + "use_bfloat16": false, + "vocab_size": 49409 + }, + "torch_dtype": "float32", + "transformers_version": null, + "vision_config": { + "_name_or_path": "", + "add_cross_attention": false, + "architectures": null, + "attention_dropout": 0.0, + "bad_words_ids": null, + "begin_suppress_tokens": null, + "bos_token_id": null, + "chunk_size_feed_forward": 0, + "cross_attention_hidden_size": null, + "decoder_start_token_id": null, + "diversity_penalty": 0.0, + "do_sample": false, + "early_stopping": false, + "encoder_no_repeat_ngram_size": 0, + "eos_token_id": null, + "exponential_decay_length_penalty": null, + "finetuning_task": null, + "forced_bos_token_id": null, + "forced_eos_token_id": null, + "hidden_act": "quick_gelu", + "hidden_size": 1280, + "id2label": { + "0": "LABEL_0", + "1": "LABEL_1" + }, + "image_size": 378, + "initializer_factor": 1.0, + "initializer_range": 0.02, + "intermediate_size": 5120, + "is_decoder": false, + "is_encoder_decoder": false, + "label2id": { + "LABEL_0": 0, + "LABEL_1": 1 + }, + "layer_norm_eps": 1e-05, + "length_penalty": 1.0, + "max_length": 20, + "min_length": 0, + "model_type": "clip_vision_model", + "no_repeat_ngram_size": 0, + "num_attention_heads": 16, + "num_beam_groups": 1, + "num_beams": 1, + "num_channels": 3, + "num_hidden_layers": 32, + "num_return_sequences": 1, + "output_attentions": false, + "output_hidden_states": false, + "output_scores": false, + "pad_token_id": null, + "patch_size": 14, + "prefix": null, + "problem_type": null, + "projection_dim": 512, + "pruned_heads": {}, + "remove_invalid_values": false, + "repetition_penalty": 1.0, + "return_dict": true, + "return_dict_in_generate": false, + "sep_token_id": null, + "suppress_tokens": null, + "task_specific_params": null, + "temperature": 1.0, + "tf_legacy_loss": false, + "tie_encoder_decoder": false, + "tie_word_embeddings": true, + "tokenizer_class": null, + "top_k": 50, + "top_p": 1.0, + "torch_dtype": null, + "torchscript": false, + "transformers_version": "4.27.1", + "typical_p": 1.0, + "use_bfloat16": false + } +} diff --git a/custom_nodes/ComfyUI-MMAudio/configs/nvidia_bigvgan_v2_config.json b/custom_nodes/ComfyUI-MMAudio/configs/nvidia_bigvgan_v2_config.json new file mode 100644 index 0000000000000000000000000000000000000000..8f02b5a20c50d4835e5f16dd881877d0904dc535 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/configs/nvidia_bigvgan_v2_config.json @@ -0,0 +1,63 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 32, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [8,4,2,2,2,2], + "upsample_kernel_sizes": [16,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "use_tanh_at_final": false, + "use_bias_at_final": false, + + "activation": "snakebeta", + "snake_logscale": true, + + "use_cqtd_instead_of_mrd": true, + "cqtd_filters": 128, + "cqtd_max_filters": 1024, + "cqtd_filters_scale": 1, + "cqtd_dilations": [1, 2, 4], + "cqtd_hop_lengths": [512, 256, 256], + "cqtd_n_octaves": [9, 9, 9], + "cqtd_bins_per_octaves": [24, 36, 48], + + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "use_multiscale_melloss": true, + "lambda_melloss": 15, + + "clip_grad_norm": 500, + + "segment_size": 65536, + "num_mels": 128, + "num_freq": 2049, + "n_fft": 2048, + "hop_size": 512, + "win_size": 2048, + + "sampling_rate": 44100, + + "fmin": 0, + "fmax": null, + "fmax_for_loss": null, + + "normalize_volume": true, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/example_workflows/mmaudio_test.json b/custom_nodes/ComfyUI-MMAudio/example_workflows/mmaudio_test.json new file mode 100644 index 0000000000000000000000000000000000000000..94ea5a3dd5719280048a3b5e7e18a0f9e0de9ba0 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/example_workflows/mmaudio_test.json @@ -0,0 +1,514 @@ +{ + "last_node_id": 106, + "last_link_id": 176, + "nodes": [ + { + "id": 92, + "type": "MMAudioSampler", + "pos": [ + 2322.641845703125, + -222.64015197753906 + ], + "size": [ + 411.29339599609375, + 354 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "name": "mmaudio_model", + "type": "MMAUDIO_MODEL", + "link": 152 + }, + { + "name": "feature_utils", + "type": "MMAUDIO_FEATUREUTILS", + "link": 166 + }, + { + "name": "images", + "type": "IMAGE", + "link": 170, + "shape": 7 + }, + { + "name": "duration", + "type": "FLOAT", + "link": 175, + "widget": { + "name": "duration" + } + } + ], + "outputs": [ + { + "name": "audio", + "type": "AUDIO", + "links": [ + 163, + 171 + ], + "slot_index": 0 + } + ], + "properties": { + "Node name for S&R": "MMAudioSampler" + }, + "widgets_values": [ + 3, + 25, + 4.5, + 7, + "fixed", + "", + "", + false, + true + ] + }, + { + "id": 98, + "type": "PreviewAudio", + "pos": [ + 2835.40087890625, + -351.83074951171875 + ], + "size": [ + 315, + 76 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "name": "audio", + "type": "AUDIO", + "link": 163 + } + ], + "outputs": [], + "properties": { + "Node name for S&R": "PreviewAudio" + }, + "widgets_values": [ + null + ] + }, + { + "id": 97, + "type": "VHS_VideoCombine", + "pos": [ + 2805.75927734375, + -169.37237548828125 + ], + "size": [ + 648.7454223632812, + 1325.99267578125 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 161 + }, + { + "name": "audio", + "type": "AUDIO", + "link": 171, + "shape": 7 + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null, + "shape": 7 + }, + { + "name": "vae", + "type": "VAE", + "link": null, + "shape": 7 + }, + { + "name": "frame_rate", + "type": "FLOAT", + "link": 176, + "widget": { + "name": "frame_rate" + } + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 24, + "loop_count": 0, + "filename_prefix": "MMaudio", + "format": "video/h264-mp4", + "pix_fmt": "yuv420p", + "crf": 19, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "MMaudio_00019-audio.mp4", + "subfolder": "", + "type": "temp", + "format": "video/h264-mp4", + "frame_rate": 16 + }, + "muted": false + } + } + }, + { + "id": 102, + "type": "MMAudioFeatureUtilsLoader", + "pos": [ + 1577.2188720703125, + -68.85494232177734 + ], + "size": [ + 487.2332763671875, + 154 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "name": "bigvgan_vocoder_model", + "type": "VOCODER_MODEL", + "link": null, + "shape": 7 + } + ], + "outputs": [ + { + "name": "mmaudio_featureutils", + "type": "MMAUDIO_FEATUREUTILS", + "links": [ + 166 + ] + } + ], + "properties": { + "Node name for S&R": "MMAudioFeatureUtilsLoader" + }, + "widgets_values": [ + "mmaudio_vae_44k_fp16.safetensors", + "mmaudio_synchformer_fp16.safetensors", + "apple_DFN5B-CLIP-ViT-H-14-384_fp16.safetensors", + "44k", + "fp16" + ] + }, + { + "id": 85, + "type": "MMAudioModelLoader", + "pos": [ + 1642.6138916015625, + -228.39370727539062 + ], + "size": [ + 438.219970703125, + 83.87998962402344 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [], + "outputs": [ + { + "name": "mmaudio_model", + "type": "MMAUDIO_MODEL", + "links": [ + 152 + ], + "slot_index": 0 + } + ], + "properties": { + "Node name for S&R": "MMAudioModelLoader" + }, + "widgets_values": [ + "mmaudio_large_44k_v2_fp16.safetensors", + "fp16" + ] + }, + { + "id": 105, + "type": "VHS_VideoInfo", + "pos": [ + 2028.3045654296875, + 236.1528778076172 + ], + "size": [ + 262, + 206 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "link": 172 + } + ], + "outputs": [ + { + "name": "source_fps🟨", + "type": "FLOAT", + "links": null + }, + { + "name": "source_frame_count🟨", + "type": "INT", + "links": null + }, + { + "name": "source_duration🟨", + "type": "FLOAT", + "links": null + }, + { + "name": "source_width🟨", + "type": "INT", + "links": null + }, + { + "name": "source_height🟨", + "type": "INT", + "links": null + }, + { + "name": "loaded_fps🟦", + "type": "FLOAT", + "links": [ + 176 + ], + "slot_index": 5 + }, + { + "name": "loaded_frame_count🟦", + "type": "INT", + "links": null, + "slot_index": 6 + }, + { + "name": "loaded_duration🟦", + "type": "FLOAT", + "links": [ + 175 + ], + "slot_index": 7 + }, + { + "name": "loaded_width🟦", + "type": "INT", + "links": null + }, + { + "name": "loaded_height🟦", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoInfo" + }, + "widgets_values": {} + }, + { + "id": 91, + "type": "VHS_LoadVideo", + "pos": [ + 1559.9124755859375, + 174.70217895507812 + ], + "size": [ + 417.6617736816406, + 912.2588500976562 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null, + "shape": 7 + }, + { + "name": "vae", + "type": "VAE", + "link": null, + "shape": 7 + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 161, + 170 + ], + "slot_index": 0 + }, + { + "name": "frame_count", + "type": "INT", + "links": null + }, + { + "name": "audio", + "type": "AUDIO", + "links": null + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": [ + 172 + ], + "slot_index": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideo" + }, + "widgets_values": { + "video": "CogVideoX_1_5_I2V_00092.mp4", + "force_rate": 0, + "force_size": "Disabled", + "custom_width": 512, + "custom_height": 512, + "frame_load_cap": 0, + "skip_first_frames": 0, + "select_every_nth": 1, + "choose video to upload": "image", + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 0, + "skip_first_frames": 0, + "force_rate": 0, + "filename": "CogVideoX_1_5_I2V_00092.mp4", + "type": "input", + "format": "video/mp4", + "select_every_nth": 1 + }, + "muted": false + } + } + } + ], + "links": [ + [ + 152, + 85, + 0, + 92, + 0, + "MMAUDIO_MODEL" + ], + [ + 161, + 91, + 0, + 97, + 0, + "IMAGE" + ], + [ + 163, + 92, + 0, + 98, + 0, + "AUDIO" + ], + [ + 166, + 102, + 0, + 92, + 1, + "MMAUDIO_FEATUREUTILS" + ], + [ + 170, + 91, + 0, + 92, + 2, + "IMAGE" + ], + [ + 171, + 92, + 0, + 97, + 1, + "AUDIO" + ], + [ + 172, + 91, + 3, + 105, + 0, + "VHS_VIDEOINFO" + ], + [ + 175, + 105, + 7, + 92, + 3, + "FLOAT" + ], + [ + 176, + 105, + 5, + 97, + 4, + "FLOAT" + ] + ], + "groups": [], + "config": {}, + "extra": { + "ds": { + "scale": 0.6830134553650706, + "offset": [ + -1061.593768518744, + 451.0850929624773 + ] + } + }, + "version": 0.4 +} \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/eval_utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/eval_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b69ab836b1996eb0237f43b6710ed20dffacb402 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/eval_utils.py @@ -0,0 +1,84 @@ +import dataclasses +import logging +from pathlib import Path +from typing import Optional + +import torch + +from .model.flow_matching import FlowMatching +from .model.networks import MMAudio +from .model.sequence_config import (CONFIG_16K, CONFIG_44K, SequenceConfig) +from .model.utils.features_utils import FeaturesUtils + +log = logging.getLogger() + +@dataclasses.dataclass +class ModelConfig: + model_name: str + model_path: Path + vae_path: Path + bigvgan_16k_path: Optional[Path] + mode: str + synchformer_ckpt: Path = Path('./ext_weights/synchformer_state_dict.pth') + + @property + def seq_cfg(self) -> SequenceConfig: + if self.mode == '16k': + return CONFIG_16K + elif self.mode == '44k': + return CONFIG_44K + +def generate(clip_video: Optional[torch.Tensor], + sync_video: Optional[torch.Tensor], + text: Optional[list[str]], + *, + negative_text: Optional[list[str]] = None, + feature_utils: FeaturesUtils, + net: MMAudio, + fm: FlowMatching, + rng: torch.Generator, + cfg_strength: float): + device = feature_utils.device + dtype = feature_utils.dtype + + bs = len(text) + if clip_video is not None: + clip_video = clip_video.to(device, dtype, non_blocking=True) + clip_features = feature_utils.encode_video_with_clip(clip_video, batch_size=bs) + else: + clip_features = net.get_empty_clip_sequence(bs) + + if sync_video is not None: + sync_video = sync_video.to(device, dtype, non_blocking=True) + sync_features = feature_utils.encode_video_with_sync(sync_video, batch_size=bs) + else: + sync_features = net.get_empty_sync_sequence(bs) + + if text is not None: + text_features = feature_utils.encode_text(text) + else: + text_features = net.get_empty_string_sequence(bs) + + if negative_text is not None: + assert len(negative_text) == bs + negative_text_features = feature_utils.encode_text(negative_text) + else: + negative_text_features = net.get_empty_string_sequence(bs) + + x0 = torch.randn(bs, + net.latent_seq_len, + net.latent_dim, + device=device, + dtype=dtype, + generator=rng) + preprocessed_conditions = net.preprocess_conditions(clip_features, sync_features, text_features) + empty_conditions = net.get_empty_conditions( + bs, negative_text_features=negative_text_features if negative_text is not None else None) + + cfg_ode_wrapper = lambda t, x: net.ode_wrapper(t, x, preprocessed_conditions, empty_conditions, + cfg_strength) + x1 = fm.to_data(cfg_ode_wrapper, x0) + x1 = net.unnormalize(x1) + spec = feature_utils.decode(x1) + audio = feature_utils.vocode(spec) + return audio diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8b137891791fe96927ad78e64b0aad7bded08bdc --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/__init__.py @@ -0,0 +1 @@ + diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e5a876391c1e48970e93ff45f212f21f86d4d0c9 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/__init__.py @@ -0,0 +1 @@ +from .autoencoder import AutoEncoderModule diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/autoencoder.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/autoencoder.py new file mode 100644 index 0000000000000000000000000000000000000000..123050dba89403a8ef378e2444d0318551240017 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/autoencoder.py @@ -0,0 +1,40 @@ +from typing import Literal, Optional + +import torch +import torch.nn as nn + +from ...ext.autoencoder.vae import VAE, get_my_vae +from ...ext.bigvgan import BigVGAN +from ...model.utils.distributions import DiagonalGaussianDistribution + + +class AutoEncoderModule(nn.Module): + + def __init__(self, + *, + vae_state_dict, + bigvgan_vocoder: Optional[BigVGAN] = None, + mode: Literal['16k', '44k']): + super().__init__() + self.vae: VAE = get_my_vae(mode).eval() + self.vae.load_state_dict(vae_state_dict) + self.vae.remove_weight_norm() + + self.vocoder = bigvgan_vocoder + if mode == '44k': + self.vocoder.remove_weight_norm() + + for param in self.parameters(): + param.requires_grad = False + + @torch.inference_mode() + def encode(self, x: torch.Tensor) -> DiagonalGaussianDistribution: + return self.vae.encode(x) + + @torch.inference_mode() + def decode(self, z: torch.Tensor) -> torch.Tensor: + return self.vae.decode(z) + + @torch.inference_mode() + def vocode(self, spec: torch.Tensor) -> torch.Tensor: + return self.vocoder(spec) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/edm2_utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/edm2_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a18ffba5cc42214fddf1300034be2eff2760025c --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/edm2_utils.py @@ -0,0 +1,168 @@ +# Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# This work is licensed under a Creative Commons +# Attribution-NonCommercial-ShareAlike 4.0 International License. +# You should have received a copy of the license along with this +# work. If not, see http://creativecommons.org/licenses/by-nc-sa/4.0/ +"""Improved diffusion model architecture proposed in the paper +"Analyzing and Improving the Training Dynamics of Diffusion Models".""" + +import numpy as np +import torch + +#---------------------------------------------------------------------------- +# Variant of constant() that inherits dtype and device from the given +# reference tensor by default. + +_constant_cache = dict() + + +def constant(value, shape=None, dtype=None, device=None, memory_format=None): + value = np.asarray(value) + if shape is not None: + shape = tuple(shape) + if dtype is None: + dtype = torch.get_default_dtype() + if device is None: + device = torch.device('cpu') + if memory_format is None: + memory_format = torch.contiguous_format + + key = (value.shape, value.dtype, value.tobytes(), shape, dtype, device, memory_format) + tensor = _constant_cache.get(key, None) + if tensor is None: + tensor = torch.as_tensor(value.copy(), dtype=dtype, device=device) + if shape is not None: + tensor, _ = torch.broadcast_tensors(tensor, torch.empty(shape)) + tensor = tensor.contiguous(memory_format=memory_format) + _constant_cache[key] = tensor + return tensor + + +def const_like(ref, value, shape=None, dtype=None, device=None, memory_format=None): + if dtype is None: + dtype = ref.dtype + if device is None: + device = ref.device + return constant(value, shape=shape, dtype=dtype, device=device, memory_format=memory_format) + + +#---------------------------------------------------------------------------- +# Normalize given tensor to unit magnitude with respect to the given +# dimensions. Default = all dimensions except the first. + + +def normalize(x, dim=None, eps=1e-4): + if dim is None: + dim = list(range(1, x.ndim)) + norm = torch.linalg.vector_norm(x, dim=dim, keepdim=True, dtype=torch.float32) + norm = torch.add(eps, norm, alpha=np.sqrt(norm.numel() / x.numel())) + return x / norm.to(x.dtype) + + +class Normalize(torch.nn.Module): + + def __init__(self, dim=None, eps=1e-4): + super().__init__() + self.dim = dim + self.eps = eps + + def forward(self, x): + return normalize(x, dim=self.dim, eps=self.eps) + + +#---------------------------------------------------------------------------- +# Upsample or downsample the given tensor with the given filter, +# or keep it as is. + + +def resample(x, f=[1, 1], mode='keep'): + if mode == 'keep': + return x + f = np.float32(f) + assert f.ndim == 1 and len(f) % 2 == 0 + pad = (len(f) - 1) // 2 + f = f / f.sum() + f = np.outer(f, f)[np.newaxis, np.newaxis, :, :] + f = const_like(x, f) + c = x.shape[1] + if mode == 'down': + return torch.nn.functional.conv2d(x, + f.tile([c, 1, 1, 1]), + groups=c, + stride=2, + padding=(pad, )) + assert mode == 'up' + return torch.nn.functional.conv_transpose2d(x, (f * 4).tile([c, 1, 1, 1]), + groups=c, + stride=2, + padding=(pad, )) + + +#---------------------------------------------------------------------------- +# Magnitude-preserving SiLU (Equation 81). + + +def mp_silu(x): + return torch.nn.functional.silu(x) / 0.596 + + +class MPSiLU(torch.nn.Module): + + def forward(self, x): + return mp_silu(x) + + +#---------------------------------------------------------------------------- +# Magnitude-preserving sum (Equation 88). + + +def mp_sum(a, b, t=0.5): + return a.lerp(b, t) / np.sqrt((1 - t)**2 + t**2) + + +#---------------------------------------------------------------------------- +# Magnitude-preserving concatenation (Equation 103). + + +def mp_cat(a, b, dim=1, t=0.5): + Na = a.shape[dim] + Nb = b.shape[dim] + C = np.sqrt((Na + Nb) / ((1 - t)**2 + t**2)) + wa = C / np.sqrt(Na) * (1 - t) + wb = C / np.sqrt(Nb) * t + return torch.cat([wa * a, wb * b], dim=dim) + + +#---------------------------------------------------------------------------- +# Magnitude-preserving convolution or fully-connected layer (Equation 47) +# with force weight normalization (Equation 66). + + +class MPConv1D(torch.nn.Module): + + def __init__(self, in_channels, out_channels, kernel_size): + super().__init__() + self.out_channels = out_channels + self.weight = torch.nn.Parameter(torch.randn(out_channels, in_channels, kernel_size)) + + self.weight_norm_removed = False + + def forward(self, x, gain=1): + assert self.weight_norm_removed, 'call remove_weight_norm() before inference' + + w = self.weight * gain + if w.ndim == 2: + return x @ w.t() + assert w.ndim == 3 + return torch.nn.functional.conv1d(x, w, padding=(w.shape[-1] // 2, )) + + def remove_weight_norm(self): + w = self.weight.to(torch.float32) + w = normalize(w) # traditional weight normalization + w = w / np.sqrt(w[0].numel()) + w = w.to(self.weight.dtype) + self.weight.data.copy_(w) + + self.weight_norm_removed = True + return self diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/vae.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/vae.py new file mode 100644 index 0000000000000000000000000000000000000000..efa373182f150f01221ad291b84b58937104709a --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/vae.py @@ -0,0 +1,369 @@ +import logging +from typing import Optional + +import torch +import torch.nn as nn + +from ...ext.autoencoder.edm2_utils import MPConv1D +from ...ext.autoencoder.vae_modules import (AttnBlock1D, Downsample1D, ResnetBlock1D, + Upsample1D, nonlinearity) +from ...model.utils.distributions import DiagonalGaussianDistribution + +log = logging.getLogger() + +DATA_MEAN_80D = [ + -1.6058, -1.3676, -1.2520, -1.2453, -1.2078, -1.2224, -1.2419, -1.2439, -1.2922, -1.2927, + -1.3170, -1.3543, -1.3401, -1.3836, -1.3907, -1.3912, -1.4313, -1.4152, -1.4527, -1.4728, + -1.4568, -1.5101, -1.5051, -1.5172, -1.5623, -1.5373, -1.5746, -1.5687, -1.6032, -1.6131, + -1.6081, -1.6331, -1.6489, -1.6489, -1.6700, -1.6738, -1.6953, -1.6969, -1.7048, -1.7280, + -1.7361, -1.7495, -1.7658, -1.7814, -1.7889, -1.8064, -1.8221, -1.8377, -1.8417, -1.8643, + -1.8857, -1.8929, -1.9173, -1.9379, -1.9531, -1.9673, -1.9824, -2.0042, -2.0215, -2.0436, + -2.0766, -2.1064, -2.1418, -2.1855, -2.2319, -2.2767, -2.3161, -2.3572, -2.3954, -2.4282, + -2.4659, -2.5072, -2.5552, -2.6074, -2.6584, -2.7107, -2.7634, -2.8266, -2.8981, -2.9673 +] + +DATA_STD_80D = [ + 1.0291, 1.0411, 1.0043, 0.9820, 0.9677, 0.9543, 0.9450, 0.9392, 0.9343, 0.9297, 0.9276, 0.9263, + 0.9242, 0.9254, 0.9232, 0.9281, 0.9263, 0.9315, 0.9274, 0.9247, 0.9277, 0.9199, 0.9188, 0.9194, + 0.9160, 0.9161, 0.9146, 0.9161, 0.9100, 0.9095, 0.9145, 0.9076, 0.9066, 0.9095, 0.9032, 0.9043, + 0.9038, 0.9011, 0.9019, 0.9010, 0.8984, 0.8983, 0.8986, 0.8961, 0.8962, 0.8978, 0.8962, 0.8973, + 0.8993, 0.8976, 0.8995, 0.9016, 0.8982, 0.8972, 0.8974, 0.8949, 0.8940, 0.8947, 0.8936, 0.8939, + 0.8951, 0.8956, 0.9017, 0.9167, 0.9436, 0.9690, 1.0003, 1.0225, 1.0381, 1.0491, 1.0545, 1.0604, + 1.0761, 1.0929, 1.1089, 1.1196, 1.1176, 1.1156, 1.1117, 1.1070 +] + +DATA_MEAN_128D = [ + -3.3462, -2.6723, -2.4893, -2.3143, -2.2664, -2.3317, -2.1802, -2.4006, -2.2357, -2.4597, + -2.3717, -2.4690, -2.5142, -2.4919, -2.6610, -2.5047, -2.7483, -2.5926, -2.7462, -2.7033, + -2.7386, -2.8112, -2.7502, -2.9594, -2.7473, -3.0035, -2.8891, -2.9922, -2.9856, -3.0157, + -3.1191, -2.9893, -3.1718, -3.0745, -3.1879, -3.2310, -3.1424, -3.2296, -3.2791, -3.2782, + -3.2756, -3.3134, -3.3509, -3.3750, -3.3951, -3.3698, -3.4505, -3.4509, -3.5089, -3.4647, + -3.5536, -3.5788, -3.5867, -3.6036, -3.6400, -3.6747, -3.7072, -3.7279, -3.7283, -3.7795, + -3.8259, -3.8447, -3.8663, -3.9182, -3.9605, -3.9861, -4.0105, -4.0373, -4.0762, -4.1121, + -4.1488, -4.1874, -4.2461, -4.3170, -4.3639, -4.4452, -4.5282, -4.6297, -4.7019, -4.7960, + -4.8700, -4.9507, -5.0303, -5.0866, -5.1634, -5.2342, -5.3242, -5.4053, -5.4927, -5.5712, + -5.6464, -5.7052, -5.7619, -5.8410, -5.9188, -6.0103, -6.0955, -6.1673, -6.2362, -6.3120, + -6.3926, -6.4797, -6.5565, -6.6511, -6.8130, -6.9961, -7.1275, -7.2457, -7.3576, -7.4663, + -7.6136, -7.7469, -7.8815, -8.0132, -8.1515, -8.3071, -8.4722, -8.7418, -9.3975, -9.6628, + -9.7671, -9.8863, -9.9992, -10.0860, -10.1709, -10.5418, -11.2795, -11.3861 +] + +DATA_STD_128D = [ + 2.3804, 2.4368, 2.3772, 2.3145, 2.2803, 2.2510, 2.2316, 2.2083, 2.1996, 2.1835, 2.1769, 2.1659, + 2.1631, 2.1618, 2.1540, 2.1606, 2.1571, 2.1567, 2.1612, 2.1579, 2.1679, 2.1683, 2.1634, 2.1557, + 2.1668, 2.1518, 2.1415, 2.1449, 2.1406, 2.1350, 2.1313, 2.1415, 2.1281, 2.1352, 2.1219, 2.1182, + 2.1327, 2.1195, 2.1137, 2.1080, 2.1179, 2.1036, 2.1087, 2.1036, 2.1015, 2.1068, 2.0975, 2.0991, + 2.0902, 2.1015, 2.0857, 2.0920, 2.0893, 2.0897, 2.0910, 2.0881, 2.0925, 2.0873, 2.0960, 2.0900, + 2.0957, 2.0958, 2.0978, 2.0936, 2.0886, 2.0905, 2.0845, 2.0855, 2.0796, 2.0840, 2.0813, 2.0817, + 2.0838, 2.0840, 2.0917, 2.1061, 2.1431, 2.1976, 2.2482, 2.3055, 2.3700, 2.4088, 2.4372, 2.4609, + 2.4731, 2.4847, 2.5072, 2.5451, 2.5772, 2.6147, 2.6529, 2.6596, 2.6645, 2.6726, 2.6803, 2.6812, + 2.6899, 2.6916, 2.6931, 2.6998, 2.7062, 2.7262, 2.7222, 2.7158, 2.7041, 2.7485, 2.7491, 2.7451, + 2.7485, 2.7233, 2.7297, 2.7233, 2.7145, 2.6958, 2.6788, 2.6439, 2.6007, 2.4786, 2.2469, 2.1877, + 2.1392, 2.0717, 2.0107, 1.9676, 1.9140, 1.7102, 0.9101, 0.7164 +] + + +class VAE(nn.Module): + + def __init__( + self, + *, + data_dim: int, + embed_dim: int, + hidden_dim: int, + ): + super().__init__() + + if data_dim == 80: + self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_80D, dtype=torch.float32).cuda()) + self.data_std = nn.Buffer(torch.tensor(DATA_STD_80D, dtype=torch.float32).cuda()) + elif data_dim == 128: + self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_128D, dtype=torch.float32).cuda()) + self.data_std = nn.Buffer(torch.tensor(DATA_STD_128D, dtype=torch.float32).cuda()) + + self.data_mean = self.data_mean.view(1, -1, 1) + self.data_std = self.data_std.view(1, -1, 1) + + self.encoder = Encoder1D( + dim=hidden_dim, + ch_mult=(1, 2, 4), + num_res_blocks=2, + attn_layers=[3], + down_layers=[0], + in_dim=data_dim, + embed_dim=embed_dim, + ) + self.decoder = Decoder1D( + dim=hidden_dim, + ch_mult=(1, 2, 4), + num_res_blocks=2, + attn_layers=[3], + down_layers=[0], + in_dim=data_dim, + out_dim=data_dim, + embed_dim=embed_dim, + ) + + self.embed_dim = embed_dim + # self.quant_conv = nn.Conv1d(2 * embed_dim, 2 * embed_dim, 1) + # self.post_quant_conv = nn.Conv1d(embed_dim, embed_dim, 1) + + self.initialize_weights() + + def initialize_weights(self): + pass + + def encode(self, x: torch.Tensor, normalize: bool = True) -> DiagonalGaussianDistribution: + if normalize: + x = self.normalize(x) + moments = self.encoder(x) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z: torch.Tensor, unnormalize: bool = True) -> torch.Tensor: + dec = self.decoder(z) + if unnormalize: + dec = self.unnormalize(dec) + return dec + + def normalize(self, x: torch.Tensor) -> torch.Tensor: + return (x - self.data_mean) / self.data_std + + def unnormalize(self, x: torch.Tensor) -> torch.Tensor: + return x * self.data_std + self.data_mean + + def forward( + self, + x: torch.Tensor, + sample_posterior: bool = True, + rng: Optional[torch.Generator] = None, + normalize: bool = True, + unnormalize: bool = True, + ) -> tuple[torch.Tensor, DiagonalGaussianDistribution]: + + posterior = self.encode(x, normalize=normalize) + if sample_posterior: + z = posterior.sample(rng) + else: + z = posterior.mode() + dec = self.decode(z, unnormalize=unnormalize) + return dec, posterior + + def load_weights(self, src_dict) -> None: + self.load_state_dict(src_dict, strict=True) + + @property + def device(self) -> torch.device: + return next(self.parameters()).device + + def get_last_layer(self): + return self.decoder.conv_out.weight + + def remove_weight_norm(self): + for name, m in self.named_modules(): + if isinstance(m, MPConv1D): + m.remove_weight_norm() + log.debug(f"Removed weight norm from {name}") + return self + + +class Encoder1D(nn.Module): + + def __init__(self, + *, + dim: int, + ch_mult: tuple[int] = (1, 2, 4, 8), + num_res_blocks: int, + attn_layers: list[int] = [], + down_layers: list[int] = [], + resamp_with_conv: bool = True, + in_dim: int, + embed_dim: int, + double_z: bool = True, + kernel_size: int = 3, + clip_act: float = 256.0): + super().__init__() + self.dim = dim + self.num_layers = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.in_channels = in_dim + self.clip_act = clip_act + self.down_layers = down_layers + self.attn_layers = attn_layers + self.conv_in = MPConv1D(in_dim, self.dim, kernel_size=kernel_size) + + in_ch_mult = (1, ) + tuple(ch_mult) + self.in_ch_mult = in_ch_mult + # downsampling + self.down = nn.ModuleList() + for i_level in range(self.num_layers): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = dim * in_ch_mult[i_level] + block_out = dim * ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append( + ResnetBlock1D(in_dim=block_in, + out_dim=block_out, + kernel_size=kernel_size, + use_norm=True)) + block_in = block_out + if i_level in attn_layers: + attn.append(AttnBlock1D(block_in)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level in down_layers: + down.downsample = Downsample1D(block_in, resamp_with_conv) + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock1D(in_dim=block_in, + out_dim=block_in, + kernel_size=kernel_size, + use_norm=True) + self.mid.attn_1 = AttnBlock1D(block_in) + self.mid.block_2 = ResnetBlock1D(in_dim=block_in, + out_dim=block_in, + kernel_size=kernel_size, + use_norm=True) + + # end + self.conv_out = MPConv1D(block_in, + 2 * embed_dim if double_z else embed_dim, + kernel_size=kernel_size) + + self.learnable_gain = nn.Parameter(torch.zeros([])) + + def forward(self, x): + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_layers): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1]) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + h = h.clamp(-self.clip_act, self.clip_act) + hs.append(h) + if i_level in self.down_layers: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h) + h = self.mid.attn_1(h) + h = self.mid.block_2(h) + h = h.clamp(-self.clip_act, self.clip_act) + + # end + h = nonlinearity(h) + h = self.conv_out(h, gain=(self.learnable_gain + 1)) + return h + + +class Decoder1D(nn.Module): + + def __init__(self, + *, + dim: int, + out_dim: int, + ch_mult: tuple[int] = (1, 2, 4, 8), + num_res_blocks: int, + attn_layers: list[int] = [], + down_layers: list[int] = [], + kernel_size: int = 3, + resamp_with_conv: bool = True, + in_dim: int, + embed_dim: int, + clip_act: float = 256.0): + super().__init__() + self.ch = dim + self.num_layers = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.in_channels = in_dim + self.clip_act = clip_act + self.down_layers = [i + 1 for i in down_layers] # each downlayer add one + + # compute in_ch_mult, block_in and curr_res at lowest res + block_in = dim * ch_mult[self.num_layers - 1] + + # z to block_in + self.conv_in = MPConv1D(embed_dim, block_in, kernel_size=kernel_size) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True) + self.mid.attn_1 = AttnBlock1D(block_in) + self.mid.block_2 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_layers)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = dim * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + block.append(ResnetBlock1D(in_dim=block_in, out_dim=block_out, use_norm=True)) + block_in = block_out + if i_level in attn_layers: + attn.append(AttnBlock1D(block_in)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level in self.down_layers: + up.upsample = Upsample1D(block_in, resamp_with_conv) + self.up.insert(0, up) # prepend to get consistent order + + # end + self.conv_out = MPConv1D(block_in, out_dim, kernel_size=kernel_size) + self.learnable_gain = nn.Parameter(torch.zeros([])) + + def forward(self, z): + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h) + h = self.mid.attn_1(h) + h = self.mid.block_2(h) + h = h.clamp(-self.clip_act, self.clip_act) + + # upsampling + for i_level in reversed(range(self.num_layers)): + for i_block in range(self.num_res_blocks + 1): + h = self.up[i_level].block[i_block](h) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + h = h.clamp(-self.clip_act, self.clip_act) + if i_level in self.down_layers: + h = self.up[i_level].upsample(h) + + h = nonlinearity(h) + h = self.conv_out(h, gain=(self.learnable_gain + 1)) + return h + + +def VAE_16k(**kwargs) -> VAE: + return VAE(data_dim=80, embed_dim=20, hidden_dim=384, **kwargs) + + +def VAE_44k(**kwargs) -> VAE: + return VAE(data_dim=128, embed_dim=40, hidden_dim=512, **kwargs) + + +def get_my_vae(name: str, **kwargs) -> VAE: + if name == '16k': + return VAE_16k(**kwargs) + if name == '44k': + return VAE_44k(**kwargs) + raise ValueError(f'Unknown model: {name}') + + +if __name__ == '__main__': + network = get_my_vae('standard') + + # print the number of parameters in terms of millions + num_params = sum(p.numel() for p in network.parameters()) / 1e6 + print(f'Number of parameters: {num_params:.2f}M') diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/vae_modules.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/vae_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..3dbd51742d7713e0203d0494866c5f6604683c30 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/autoencoder/vae_modules.py @@ -0,0 +1,117 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange + +from ...ext.autoencoder.edm2_utils import (MPConv1D, mp_silu, mp_sum, normalize) + + +def nonlinearity(x): + # swish + return mp_silu(x) + + +class ResnetBlock1D(nn.Module): + + def __init__(self, *, in_dim, out_dim=None, conv_shortcut=False, kernel_size=3, use_norm=True): + super().__init__() + self.in_dim = in_dim + out_dim = in_dim if out_dim is None else out_dim + self.out_dim = out_dim + self.use_conv_shortcut = conv_shortcut + self.use_norm = use_norm + + self.conv1 = MPConv1D(in_dim, out_dim, kernel_size=kernel_size) + self.conv2 = MPConv1D(out_dim, out_dim, kernel_size=kernel_size) + if self.in_dim != self.out_dim: + if self.use_conv_shortcut: + self.conv_shortcut = MPConv1D(in_dim, out_dim, kernel_size=kernel_size) + else: + self.nin_shortcut = MPConv1D(in_dim, out_dim, kernel_size=1) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + + # pixel norm + if self.use_norm: + x = normalize(x, dim=1) + + h = x + h = nonlinearity(h) + h = self.conv1(h) + + h = nonlinearity(h) + h = self.conv2(h) + + if self.in_dim != self.out_dim: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return mp_sum(x, h, t=0.3) + + +class AttnBlock1D(nn.Module): + + def __init__(self, in_channels, num_heads=1): + super().__init__() + self.in_channels = in_channels + + self.num_heads = num_heads + self.qkv = MPConv1D(in_channels, in_channels * 3, kernel_size=1) + self.proj_out = MPConv1D(in_channels, in_channels, kernel_size=1) + + def forward(self, x): + h = x + y = self.qkv(h) + y = y.reshape(y.shape[0], self.num_heads, -1, 3, y.shape[-1]) + q, k, v = normalize(y, dim=2).unbind(3) + + q = rearrange(q, 'b h c l -> b h l c') + k = rearrange(k, 'b h c l -> b h l c') + v = rearrange(v, 'b h c l -> b h l c') + + h = F.scaled_dot_product_attention(q, k, v) + h = rearrange(h, 'b h l c -> b (h c) l') + + h = self.proj_out(h) + + return mp_sum(x, h, t=0.3) + + +class Upsample1D(nn.Module): + + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = MPConv1D(in_channels, in_channels, kernel_size=3) + + def forward(self, x): + x = F.interpolate(x, scale_factor=2.0, mode='nearest-exact') # support 3D tensor(B,C,T) + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample1D(nn.Module): + + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv1 = MPConv1D(in_channels, in_channels, kernel_size=1) + self.conv2 = MPConv1D(in_channels, in_channels, kernel_size=1) + + def forward(self, x): + + if self.with_conv: + x = self.conv1(x) + + x = F.avg_pool1d(x, kernel_size=2, stride=2) + + if self.with_conv: + x = self.conv2(x) + + return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/LICENSE b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..e9663595cc28938f88d6299acd3ba791542e4c0c --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2022 NVIDIA CORPORATION. + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..00f13e9bf9ccb0b4ec37e1c70869f9a9a538871f --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/__init__.py @@ -0,0 +1 @@ +from .bigvgan import BigVGAN diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/activations.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/activations.py new file mode 100644 index 0000000000000000000000000000000000000000..61f2808a5466b3cf4d041059700993af5527dd29 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/activations.py @@ -0,0 +1,120 @@ +# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license. +# LICENSE is in incl_licenses directory. + +import torch +from torch import nn, sin, pow +from torch.nn import Parameter + + +class Snake(nn.Module): + ''' + Implementation of a sine-based periodic activation function + Shape: + - Input: (B, C, T) + - Output: (B, C, T), same shape as the input + Parameters: + - alpha - trainable parameter + References: + - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda: + https://arxiv.org/abs/2006.08195 + Examples: + >>> a1 = snake(256) + >>> x = torch.randn(256) + >>> x = a1(x) + ''' + def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False): + ''' + Initialization. + INPUT: + - in_features: shape of the input + - alpha: trainable parameter + alpha is initialized to 1 by default, higher values = higher-frequency. + alpha will be trained along with the rest of your model. + ''' + super(Snake, self).__init__() + self.in_features = in_features + + # initialize alpha + self.alpha_logscale = alpha_logscale + if self.alpha_logscale: # log scale alphas initialized to zeros + self.alpha = Parameter(torch.zeros(in_features) * alpha) + else: # linear scale alphas initialized to ones + self.alpha = Parameter(torch.ones(in_features) * alpha) + + self.alpha.requires_grad = alpha_trainable + + self.no_div_by_zero = 0.000000001 + + def forward(self, x): + ''' + Forward pass of the function. + Applies the function to the input elementwise. + Snake ∶= x + 1/a * sin^2 (xa) + ''' + alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T] + if self.alpha_logscale: + alpha = torch.exp(alpha) + x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2) + + return x + + +class SnakeBeta(nn.Module): + ''' + A modified Snake function which uses separate parameters for the magnitude of the periodic components + Shape: + - Input: (B, C, T) + - Output: (B, C, T), same shape as the input + Parameters: + - alpha - trainable parameter that controls frequency + - beta - trainable parameter that controls magnitude + References: + - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda: + https://arxiv.org/abs/2006.08195 + Examples: + >>> a1 = snakebeta(256) + >>> x = torch.randn(256) + >>> x = a1(x) + ''' + def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False): + ''' + Initialization. + INPUT: + - in_features: shape of the input + - alpha - trainable parameter that controls frequency + - beta - trainable parameter that controls magnitude + alpha is initialized to 1 by default, higher values = higher-frequency. + beta is initialized to 1 by default, higher values = higher-magnitude. + alpha will be trained along with the rest of your model. + ''' + super(SnakeBeta, self).__init__() + self.in_features = in_features + + # initialize alpha + self.alpha_logscale = alpha_logscale + if self.alpha_logscale: # log scale alphas initialized to zeros + self.alpha = Parameter(torch.zeros(in_features) * alpha) + self.beta = Parameter(torch.zeros(in_features) * alpha) + else: # linear scale alphas initialized to ones + self.alpha = Parameter(torch.ones(in_features) * alpha) + self.beta = Parameter(torch.ones(in_features) * alpha) + + self.alpha.requires_grad = alpha_trainable + self.beta.requires_grad = alpha_trainable + + self.no_div_by_zero = 0.000000001 + + def forward(self, x): + ''' + Forward pass of the function. + Applies the function to the input elementwise. + SnakeBeta ∶= x + 1/b * sin^2 (xa) + ''' + alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T] + beta = self.beta.unsqueeze(0).unsqueeze(-1) + if self.alpha_logscale: + alpha = torch.exp(alpha) + beta = torch.exp(beta) + x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2) + + return x \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a2318b63198250856809c0cb46210a4147b829bc --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/__init__.py @@ -0,0 +1,6 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +from .filter import * +from .resample import * +from .act import * \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/act.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/act.py new file mode 100644 index 0000000000000000000000000000000000000000..028debd697dd60458aae75010057df038bd3518a --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/act.py @@ -0,0 +1,28 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch.nn as nn +from .resample import UpSample1d, DownSample1d + + +class Activation1d(nn.Module): + def __init__(self, + activation, + up_ratio: int = 2, + down_ratio: int = 2, + up_kernel_size: int = 12, + down_kernel_size: int = 12): + super().__init__() + self.up_ratio = up_ratio + self.down_ratio = down_ratio + self.act = activation + self.upsample = UpSample1d(up_ratio, up_kernel_size) + self.downsample = DownSample1d(down_ratio, down_kernel_size) + + # x: [B,C,T] + def forward(self, x): + x = self.upsample(x) + x = self.act(x) + x = self.downsample(x) + + return x \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/filter.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/filter.py new file mode 100644 index 0000000000000000000000000000000000000000..7ad6ea87c1f10ddd94c544037791d7a4634d5ae1 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/filter.py @@ -0,0 +1,95 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch +import torch.nn as nn +import torch.nn.functional as F +import math + +if 'sinc' in dir(torch): + sinc = torch.sinc +else: + # This code is adopted from adefossez's julius.core.sinc under the MIT License + # https://adefossez.github.io/julius/julius/core.html + # LICENSE is in incl_licenses directory. + def sinc(x: torch.Tensor): + """ + Implementation of sinc, i.e. sin(pi * x) / (pi * x) + __Warning__: Different to julius.sinc, the input is multiplied by `pi`! + """ + return torch.where(x == 0, + torch.tensor(1., device=x.device, dtype=x.dtype), + torch.sin(math.pi * x) / math.pi / x) + + +# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License +# https://adefossez.github.io/julius/julius/lowpass.html +# LICENSE is in incl_licenses directory. +def kaiser_sinc_filter1d(cutoff, half_width, kernel_size): # return filter [1,1,kernel_size] + even = (kernel_size % 2 == 0) + half_size = kernel_size // 2 + + #For kaiser window + delta_f = 4 * half_width + A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95 + if A > 50.: + beta = 0.1102 * (A - 8.7) + elif A >= 21.: + beta = 0.5842 * (A - 21)**0.4 + 0.07886 * (A - 21.) + else: + beta = 0. + window = torch.kaiser_window(kernel_size, beta=beta, periodic=False) + + # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio + if even: + time = (torch.arange(-half_size, half_size) + 0.5) + else: + time = torch.arange(kernel_size) - half_size + if cutoff == 0: + filter_ = torch.zeros_like(time) + else: + filter_ = 2 * cutoff * window * sinc(2 * cutoff * time) + # Normalize filter to have sum = 1, otherwise we will have a small leakage + # of the constant component in the input signal. + filter_ /= filter_.sum() + filter = filter_.view(1, 1, kernel_size) + + return filter + + +class LowPassFilter1d(nn.Module): + def __init__(self, + cutoff=0.5, + half_width=0.6, + stride: int = 1, + padding: bool = True, + padding_mode: str = 'replicate', + kernel_size: int = 12): + # kernel_size should be even number for stylegan3 setup, + # in this implementation, odd number is also possible. + super().__init__() + if cutoff < -0.: + raise ValueError("Minimum cutoff must be larger than zero.") + if cutoff > 0.5: + raise ValueError("A cutoff above 0.5 does not make sense.") + self.kernel_size = kernel_size + self.even = (kernel_size % 2 == 0) + self.pad_left = kernel_size // 2 - int(self.even) + self.pad_right = kernel_size // 2 + self.stride = stride + self.padding = padding + self.padding_mode = padding_mode + filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size) + self.register_buffer("filter", filter) + + #input [B, C, T] + def forward(self, x): + _, C, _ = x.shape + + if self.padding: + x = F.pad(x, (self.pad_left, self.pad_right), + mode=self.padding_mode) + out = F.conv1d(x, self.filter.expand(C, -1, -1), + stride=self.stride, groups=C) + + return out \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/resample.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/resample.py new file mode 100644 index 0000000000000000000000000000000000000000..750e6c3402cc5ac939c4b9d075246562e0e1d1a7 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/alias_free_torch/resample.py @@ -0,0 +1,49 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch.nn as nn +from torch.nn import functional as F +from .filter import LowPassFilter1d +from .filter import kaiser_sinc_filter1d + + +class UpSample1d(nn.Module): + def __init__(self, ratio=2, kernel_size=None): + super().__init__() + self.ratio = ratio + self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size + self.stride = ratio + self.pad = self.kernel_size // ratio - 1 + self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2 + self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2 + filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio, + half_width=0.6 / ratio, + kernel_size=self.kernel_size) + self.register_buffer("filter", filter) + + # x: [B, C, T] + def forward(self, x): + _, C, _ = x.shape + + x = F.pad(x, (self.pad, self.pad), mode='replicate') + x = self.ratio * F.conv_transpose1d( + x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C) + x = x[..., self.pad_left:-self.pad_right] + + return x + + +class DownSample1d(nn.Module): + def __init__(self, ratio=2, kernel_size=None): + super().__init__() + self.ratio = ratio + self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size + self.lowpass = LowPassFilter1d(cutoff=0.5 / ratio, + half_width=0.6 / ratio, + stride=ratio, + kernel_size=self.kernel_size) + + def forward(self, x): + xx = self.lowpass(x) + + return xx \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/bigvgan.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/bigvgan.py new file mode 100644 index 0000000000000000000000000000000000000000..7a72a204c5b8cff45c08d0495990319a339d0937 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/bigvgan.py @@ -0,0 +1,36 @@ +from pathlib import Path + +import torch +import torch.nn as nn +from omegaconf import OmegaConf + +from ...ext.bigvgan.models import BigVGANVocoder + +_bigvgan_vocoder_path = Path(__file__).parent / 'bigvgan_vocoder.yml' + +class BigVGAN(nn.Module): + + def __init__(self, ckpt_path, config_path=_bigvgan_vocoder_path): + super().__init__() + vocoder_cfg = OmegaConf.load(config_path) + self.vocoder = BigVGANVocoder(vocoder_cfg).eval() + vocoder_ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=True)['generator'] + self.vocoder.load_state_dict(vocoder_ckpt) + + self.weight_norm_removed = False + self.remove_weight_norm() + + @classmethod + def from_pretrained(cls, ckpt_path, config_path=_bigvgan_vocoder_path): + # Create an instance of BigVGAN using the provided checkpoint path + return cls(ckpt_path, config_path) + + @torch.inference_mode() + def forward(self, x): + assert self.weight_norm_removed, 'call remove_weight_norm() before inference' + return self.vocoder(x) + + def remove_weight_norm(self): + self.vocoder.remove_weight_norm() + self.weight_norm_removed = True + return self diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/bigvgan_vocoder.yml b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/bigvgan_vocoder.yml new file mode 100644 index 0000000000000000000000000000000000000000..d4db31ec45336e757d94d5099ed16cb3c906c24a --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/bigvgan_vocoder.yml @@ -0,0 +1,63 @@ +resblock: '1' +num_gpus: 0 +batch_size: 64 +num_mels: 80 +learning_rate: 0.0001 +adam_b1: 0.8 +adam_b2: 0.99 +lr_decay: 0.999 +seed: 1234 +upsample_rates: +- 4 +- 4 +- 2 +- 2 +- 2 +- 2 +upsample_kernel_sizes: +- 8 +- 8 +- 4 +- 4 +- 4 +- 4 +upsample_initial_channel: 1536 +resblock_kernel_sizes: +- 3 +- 7 +- 11 +resblock_dilation_sizes: +- - 1 + - 3 + - 5 +- - 1 + - 3 + - 5 +- - 1 + - 3 + - 5 +activation: snakebeta +snake_logscale: true +resolutions: +- - 1024 + - 120 + - 600 +- - 2048 + - 240 + - 1200 +- - 512 + - 50 + - 240 +mpd_reshapes: +- 2 +- 3 +- 5 +- 7 +- 11 +use_spectral_norm: false +discriminator_channel_mult: 1 +num_workers: 4 +dist_config: + dist_backend: nccl + dist_url: tcp://localhost:54341 + world_size: 1 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/env.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/env.py new file mode 100644 index 0000000000000000000000000000000000000000..b8be238d4db710c8c9a338d336baea0138f18d1f --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/env.py @@ -0,0 +1,18 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import os +import shutil + + +class AttrDict(dict): + def __init__(self, *args, **kwargs): + super(AttrDict, self).__init__(*args, **kwargs) + self.__dict__ = self + + +def build_env(config, config_name, path): + t_path = os.path.join(path, config_name) + if config != t_path: + os.makedirs(path, exist_ok=True) + shutil.copyfile(config, os.path.join(path, config_name)) \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_1 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_1 new file mode 100644 index 0000000000000000000000000000000000000000..5afae394d6b37da0e12ba6b290d2512687f421ac --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_1 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Jungil Kong + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_2 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_2 new file mode 100644 index 0000000000000000000000000000000000000000..322b758863c4219be68291ae3826218baa93cb4c --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_2 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Edward Dixon + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_3 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_3 new file mode 100644 index 0000000000000000000000000000000000000000..56ee3c8c4cc2b4b32e0975d17258f9ba515fdbcc --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_3 @@ -0,0 +1,201 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_4 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_4 new file mode 100644 index 0000000000000000000000000000000000000000..48fd1a1ba8d81a94b6c7d1c2ff1a1f307cc5371d --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_4 @@ -0,0 +1,29 @@ +BSD 3-Clause License + +Copyright (c) 2019, Seungwon Park 박승원 +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +3. Neither the name of the copyright holder nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_5 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_5 new file mode 100644 index 0000000000000000000000000000000000000000..01ae5538e6b7c787bb4f5d6f2cd9903520d6e465 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/incl_licenses/LICENSE_5 @@ -0,0 +1,16 @@ +Copyright 2020 Alexandre Défossez + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and +associated documentation files (the "Software"), to deal in the Software without restriction, +including without limitation the rights to use, copy, modify, merge, publish, distribute, +sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or +substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT +NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND +NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, +DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/models.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/models.py new file mode 100644 index 0000000000000000000000000000000000000000..3e2b7d64ee04ea79401a7f6af7bcb9379661c0c3 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/models.py @@ -0,0 +1,255 @@ +# Copyright (c) 2022 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import torch +import torch.nn as nn +from torch.nn import Conv1d, ConvTranspose1d +from torch.nn.utils.parametrizations import weight_norm +from torch.nn.utils.parametrize import remove_parametrizations + +from ...ext.bigvgan import activations +from ...ext.bigvgan.alias_free_torch import * +from ...ext.bigvgan.utils import get_padding, init_weights + +LRELU_SLOPE = 0.1 + + +class AMPBlock1(torch.nn.Module): + + def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5), activation=None): + super(AMPBlock1, self).__init__() + self.h = h + + self.convs1 = nn.ModuleList([ + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]))), + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]))), + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=dilation[2], + padding=get_padding(kernel_size, dilation[2]))) + ]) + self.convs1.apply(init_weights) + + self.convs2 = nn.ModuleList([ + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1))), + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1))), + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1))) + ]) + self.convs2.apply(init_weights) + + self.num_layers = len(self.convs1) + len(self.convs2) # total number of conv layers + + if activation == 'snake': # periodic nonlinearity with snake function and anti-aliasing + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.Snake(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + elif activation == 'snakebeta': # periodic nonlinearity with snakebeta function and anti-aliasing + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + def forward(self, x): + acts1, acts2 = self.activations[::2], self.activations[1::2] + for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2): + xt = a1(x) + xt = c1(xt) + xt = a2(xt) + xt = c2(xt) + x = xt + x + + return x + + def remove_weight_norm(self): + for l in self.convs1: + remove_parametrizations(l, 'weight') + for l in self.convs2: + remove_parametrizations(l, 'weight') + + +class AMPBlock2(torch.nn.Module): + + def __init__(self, h, channels, kernel_size=3, dilation=(1, 3), activation=None): + super(AMPBlock2, self).__init__() + self.h = h + + self.convs = nn.ModuleList([ + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]))), + weight_norm( + Conv1d(channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]))) + ]) + self.convs.apply(init_weights) + + self.num_layers = len(self.convs) # total number of conv layers + + if activation == 'snake': # periodic nonlinearity with snake function and anti-aliasing + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.Snake(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + elif activation == 'snakebeta': # periodic nonlinearity with snakebeta function and anti-aliasing + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + def forward(self, x): + for c, a in zip(self.convs, self.activations): + xt = a(x) + xt = c(xt) + x = xt + x + + return x + + def remove_weight_norm(self): + for l in self.convs: + remove_parametrizations(l, 'weight') + + +class BigVGANVocoder(torch.nn.Module): + # this is our main BigVGAN model. Applies anti-aliased periodic activation for resblocks. + def __init__(self, h): + super().__init__() + self.h = h + + self.num_kernels = len(h.resblock_kernel_sizes) + self.num_upsamples = len(h.upsample_rates) + + # pre conv + self.conv_pre = weight_norm(Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)) + + # define which AMPBlock to use. BigVGAN uses AMPBlock1 as default + resblock = AMPBlock1 if h.resblock == '1' else AMPBlock2 + + # transposed conv-based upsamplers. does not apply anti-aliasing + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)): + self.ups.append( + nn.ModuleList([ + weight_norm( + ConvTranspose1d(h.upsample_initial_channel // (2**i), + h.upsample_initial_channel // (2**(i + 1)), + k, + u, + padding=(k - u) // 2)) + ])) + + # residual blocks using anti-aliased multi-periodicity composition modules (AMP) + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = h.upsample_initial_channel // (2**(i + 1)) + for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)): + self.resblocks.append(resblock(h, ch, k, d, activation=h.activation)) + + # post conv + if h.activation == "snake": # periodic nonlinearity with snake function and anti-aliasing + activation_post = activations.Snake(ch, alpha_logscale=h.snake_logscale) + self.activation_post = Activation1d(activation=activation_post) + elif h.activation == "snakebeta": # periodic nonlinearity with snakebeta function and anti-aliasing + activation_post = activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale) + self.activation_post = Activation1d(activation=activation_post) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3)) + + # weight initialization + for i in range(len(self.ups)): + self.ups[i].apply(init_weights) + self.conv_post.apply(init_weights) + + def forward(self, x): + # pre conv + x = self.conv_pre(x) + + for i in range(self.num_upsamples): + # upsampling + for i_up in range(len(self.ups[i])): + x = self.ups[i][i_up](x) + # AMP blocks + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + + # post conv + x = self.activation_post(x) + x = self.conv_post(x) + x = torch.tanh(x) + + return x + + def remove_weight_norm(self): + print('Removing weight norm...') + for l in self.ups: + for l_i in l: + remove_parametrizations(l_i, 'weight') + for l in self.resblocks: + l.remove_weight_norm() + remove_parametrizations(self.conv_pre, 'weight') + remove_parametrizations(self.conv_post, 'weight') diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..aff7e653533d3390756c53a0215801b06cc924b5 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan/utils.py @@ -0,0 +1,31 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import os + +import torch +from torch.nn.utils.parametrizations import weight_norm + + +def init_weights(m, mean=0.0, std=0.01): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + m.weight.data.normal_(mean, std) + + +def apply_weight_norm(m): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + weight_norm(m) + + +def get_padding(kernel_size, dilation=1): + return int((kernel_size * dilation - dilation) / 2) + + +def load_checkpoint(filepath, device): + assert os.path.isfile(filepath) + print("Loading '{}'".format(filepath)) + checkpoint_dict = torch.load(filepath, map_location=device) + print("Complete.") + return checkpoint_dict diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/LICENSE b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..4c78361c86d4f685117d60d6623e2197fcfed706 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2024 NVIDIA CORPORATION. + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/activations.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/activations.py new file mode 100644 index 0000000000000000000000000000000000000000..4f08ddab5b55d6dcaf3e968af98889e0770c44f5 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/activations.py @@ -0,0 +1,126 @@ +# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license. +# LICENSE is in incl_licenses directory. + +import torch +from torch import nn, sin, pow +from torch.nn import Parameter + + +class Snake(nn.Module): + """ + Implementation of a sine-based periodic activation function + Shape: + - Input: (B, C, T) + - Output: (B, C, T), same shape as the input + Parameters: + - alpha - trainable parameter + References: + - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda: + https://arxiv.org/abs/2006.08195 + Examples: + >>> a1 = snake(256) + >>> x = torch.randn(256) + >>> x = a1(x) + """ + + def __init__( + self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False + ): + """ + Initialization. + INPUT: + - in_features: shape of the input + - alpha: trainable parameter + alpha is initialized to 1 by default, higher values = higher-frequency. + alpha will be trained along with the rest of your model. + """ + super(Snake, self).__init__() + self.in_features = in_features + + # Initialize alpha + self.alpha_logscale = alpha_logscale + if self.alpha_logscale: # Log scale alphas initialized to zeros + self.alpha = Parameter(torch.zeros(in_features) * alpha) + else: # Linear scale alphas initialized to ones + self.alpha = Parameter(torch.ones(in_features) * alpha) + + self.alpha.requires_grad = alpha_trainable + + self.no_div_by_zero = 0.000000001 + + def forward(self, x): + """ + Forward pass of the function. + Applies the function to the input elementwise. + Snake ∶= x + 1/a * sin^2 (xa) + """ + alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # Line up with x to [B, C, T] + if self.alpha_logscale: + alpha = torch.exp(alpha) + x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2) + + return x + + +class SnakeBeta(nn.Module): + """ + A modified Snake function which uses separate parameters for the magnitude of the periodic components + Shape: + - Input: (B, C, T) + - Output: (B, C, T), same shape as the input + Parameters: + - alpha - trainable parameter that controls frequency + - beta - trainable parameter that controls magnitude + References: + - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda: + https://arxiv.org/abs/2006.08195 + Examples: + >>> a1 = snakebeta(256) + >>> x = torch.randn(256) + >>> x = a1(x) + """ + + def __init__( + self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False + ): + """ + Initialization. + INPUT: + - in_features: shape of the input + - alpha - trainable parameter that controls frequency + - beta - trainable parameter that controls magnitude + alpha is initialized to 1 by default, higher values = higher-frequency. + beta is initialized to 1 by default, higher values = higher-magnitude. + alpha will be trained along with the rest of your model. + """ + super(SnakeBeta, self).__init__() + self.in_features = in_features + + # Initialize alpha + self.alpha_logscale = alpha_logscale + if self.alpha_logscale: # Log scale alphas initialized to zeros + self.alpha = Parameter(torch.zeros(in_features) * alpha) + self.beta = Parameter(torch.zeros(in_features) * alpha) + else: # Linear scale alphas initialized to ones + self.alpha = Parameter(torch.ones(in_features) * alpha) + self.beta = Parameter(torch.ones(in_features) * alpha) + + self.alpha.requires_grad = alpha_trainable + self.beta.requires_grad = alpha_trainable + + self.no_div_by_zero = 0.000000001 + + def forward(self, x): + """ + Forward pass of the function. + Applies the function to the input elementwise. + SnakeBeta ∶= x + 1/b * sin^2 (xa) + """ + alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # Line up with x to [B, C, T] + beta = self.beta.unsqueeze(0).unsqueeze(-1) + if self.alpha_logscale: + alpha = torch.exp(alpha) + beta = torch.exp(beta) + x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2) + + return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/activation1d.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/activation1d.py new file mode 100644 index 0000000000000000000000000000000000000000..fbc0fd8f28a37ad949fbdb9832f51b5b933c6ff2 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/activation1d.py @@ -0,0 +1,77 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import torch +import torch.nn as nn +from alias_free_activation.torch.resample import UpSample1d, DownSample1d + +# load fused CUDA kernel: this enables importing anti_alias_activation_cuda +from alias_free_activation.cuda import load + +anti_alias_activation_cuda = load.load() + + +class FusedAntiAliasActivation(torch.autograd.Function): + """ + Assumes filter size 12, replication padding on upsampling/downsampling, and logscale alpha/beta parameters as inputs. + The hyperparameters are hard-coded in the kernel to maximize speed. + NOTE: The fused kenrel is incorrect for Activation1d with different hyperparameters. + """ + + @staticmethod + def forward(ctx, inputs, up_ftr, down_ftr, alpha, beta): + activation_results = anti_alias_activation_cuda.forward( + inputs, up_ftr, down_ftr, alpha, beta + ) + + return activation_results + + @staticmethod + def backward(ctx, output_grads): + raise NotImplementedError + return output_grads, None, None + + +class Activation1d(nn.Module): + def __init__( + self, + activation, + up_ratio: int = 2, + down_ratio: int = 2, + up_kernel_size: int = 12, + down_kernel_size: int = 12, + fused: bool = True, + ): + super().__init__() + self.up_ratio = up_ratio + self.down_ratio = down_ratio + self.act = activation + self.upsample = UpSample1d(up_ratio, up_kernel_size) + self.downsample = DownSample1d(down_ratio, down_kernel_size) + + self.fused = fused # Whether to use fused CUDA kernel or not + + def forward(self, x): + if not self.fused: + x = self.upsample(x) + x = self.act(x) + x = self.downsample(x) + return x + else: + if self.act.__class__.__name__ == "Snake": + beta = self.act.alpha.data # Snake uses same params for alpha and beta + else: + beta = ( + self.act.beta.data + ) # Snakebeta uses different params for alpha and beta + alpha = self.act.alpha.data + if ( + not self.act.alpha_logscale + ): # Exp baked into cuda kernel, cancel it out with a log + alpha = torch.log(alpha) + beta = torch.log(beta) + + x = FusedAntiAliasActivation.apply( + x, self.upsample.filter, self.downsample.lowpass.filter, alpha, beta + ) + return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation.cpp b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c5651f77143bd678169eb11564a7cf7a7969a59e --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation.cpp @@ -0,0 +1,23 @@ +/* coding=utf-8 + * Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + #include + +extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta); + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &fwd_cuda, "Anti-Alias Activation forward (CUDA)"); +} \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation_cuda.cu b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation_cuda.cu new file mode 100644 index 0000000000000000000000000000000000000000..8c442334869fe72d639ec203fa4fac07f96a0ee1 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation_cuda.cu @@ -0,0 +1,246 @@ +/* coding=utf-8 + * Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include +#include +#include +#include +#include +#include +#include +#include "type_shim.h" +#include +#include +#include +#include +#include + +namespace +{ + // Hard-coded hyperparameters + // WARP_SIZE and WARP_BATCH must match the return values batches_per_warp and + constexpr int ELEMENTS_PER_LDG_STG = 1; //(WARP_ITERATIONS < 4) ? 1 : 4; + constexpr int BUFFER_SIZE = 32; + constexpr int FILTER_SIZE = 12; + constexpr int HALF_FILTER_SIZE = 6; + constexpr int UPSAMPLE_REPLICATION_PAD = 5; // 5 on each side, matching torch impl + constexpr int DOWNSAMPLE_REPLICATION_PAD_LEFT = 5; // matching torch impl + constexpr int DOWNSAMPLE_REPLICATION_PAD_RIGHT = 6; // matching torch impl + + template + __global__ void anti_alias_activation_forward( + output_t *dst, + const input_t *src, + const input_t *up_ftr, + const input_t *down_ftr, + const input_t *alpha, + const input_t *beta, + int batch_size, + int channels, + int seq_len) + { + // Up and downsample filters + input_t up_filter[FILTER_SIZE]; + input_t down_filter[FILTER_SIZE]; + + // Load data from global memory including extra indices reserved for replication paddings + input_t elements[2 * FILTER_SIZE + 2 * BUFFER_SIZE + 2 * UPSAMPLE_REPLICATION_PAD] = {0}; + input_t intermediates[2 * FILTER_SIZE + 2 * BUFFER_SIZE + DOWNSAMPLE_REPLICATION_PAD_LEFT + DOWNSAMPLE_REPLICATION_PAD_RIGHT] = {0}; + + // Output stores downsampled output before writing to dst + output_t output[BUFFER_SIZE]; + + // blockDim/threadIdx = (128, 1, 1) + // gridDim/blockIdx = (seq_blocks, channels, batches) + int block_offset = (blockIdx.x * 128 * BUFFER_SIZE + seq_len * (blockIdx.y + gridDim.y * blockIdx.z)); + int local_offset = threadIdx.x * BUFFER_SIZE; + int seq_offset = blockIdx.x * 128 * BUFFER_SIZE + local_offset; + + // intermediate have double the seq_len + int intermediate_local_offset = threadIdx.x * BUFFER_SIZE * 2; + int intermediate_seq_offset = blockIdx.x * 128 * BUFFER_SIZE * 2 + intermediate_local_offset; + + // Get values needed for replication padding before moving pointer + const input_t *right_most_pntr = src + (seq_len * (blockIdx.y + gridDim.y * blockIdx.z)); + input_t seq_left_most_value = right_most_pntr[0]; + input_t seq_right_most_value = right_most_pntr[seq_len - 1]; + + // Move src and dst pointers + src += block_offset + local_offset; + dst += block_offset + local_offset; + + // Alpha and beta values for snake activatons. Applies exp by default + alpha = alpha + blockIdx.y; + input_t alpha_val = expf(alpha[0]); + beta = beta + blockIdx.y; + input_t beta_val = expf(beta[0]); + + #pragma unroll + for (int it = 0; it < FILTER_SIZE; it += 1) + { + up_filter[it] = up_ftr[it]; + down_filter[it] = down_ftr[it]; + } + + // Apply replication padding for upsampling, matching torch impl + #pragma unroll + for (int it = -HALF_FILTER_SIZE; it < BUFFER_SIZE + HALF_FILTER_SIZE; it += 1) + { + int element_index = seq_offset + it; // index for element + if ((element_index < 0) && (element_index >= -UPSAMPLE_REPLICATION_PAD)) + { + elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_left_most_value; + } + if ((element_index >= seq_len) && (element_index < seq_len + UPSAMPLE_REPLICATION_PAD)) + { + elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_right_most_value; + } + if ((element_index >= 0) && (element_index < seq_len)) + { + elements[2 * (HALF_FILTER_SIZE + it)] = 2 * src[it]; + } + } + + // Apply upsampling strided convolution and write to intermediates. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT for replication padding of the downsampilng conv later + #pragma unroll + for (int it = 0; it < (2 * BUFFER_SIZE + 2 * FILTER_SIZE); it += 1) + { + input_t acc = 0.0; + int element_index = intermediate_seq_offset + it; // index for intermediate + #pragma unroll + for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1) + { + if ((element_index + f_idx) >= 0) + { + acc += up_filter[f_idx] * elements[it + f_idx]; + } + } + intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] = acc; + } + + // Apply activation function. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT and DOWNSAMPLE_REPLICATION_PAD_RIGHT for replication padding of the downsampilng conv later + double no_div_by_zero = 0.000000001; + #pragma unroll + for (int it = 0; it < 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it += 1) + { + intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] += (1.0 / (beta_val + no_div_by_zero)) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val); + } + + // Apply replication padding before downsampling conv from intermediates + #pragma unroll + for (int it = 0; it < DOWNSAMPLE_REPLICATION_PAD_LEFT; it += 1) + { + intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT]; + } + #pragma unroll + for (int it = DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it < DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE + DOWNSAMPLE_REPLICATION_PAD_RIGHT; it += 1) + { + intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE - 1]; + } + + // Apply downsample strided convolution (assuming stride=2) from intermediates + #pragma unroll + for (int it = 0; it < BUFFER_SIZE; it += 1) + { + input_t acc = 0.0; + #pragma unroll + for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1) + { + // Add constant DOWNSAMPLE_REPLICATION_PAD_RIGHT to match torch implementation + acc += down_filter[f_idx] * intermediates[it * 2 + f_idx + DOWNSAMPLE_REPLICATION_PAD_RIGHT]; + } + output[it] = acc; + } + + // Write output to dst + #pragma unroll + for (int it = 0; it < BUFFER_SIZE; it += ELEMENTS_PER_LDG_STG) + { + int element_index = seq_offset + it; + if (element_index < seq_len) + { + dst[it] = output[it]; + } + } + + } + + template + void dispatch_anti_alias_activation_forward( + output_t *dst, + const input_t *src, + const input_t *up_ftr, + const input_t *down_ftr, + const input_t *alpha, + const input_t *beta, + int batch_size, + int channels, + int seq_len) + { + if (seq_len == 0) + { + return; + } + else + { + // Use 128 threads per block to maximimize gpu utilization + constexpr int threads_per_block = 128; + constexpr int seq_len_per_block = 4096; + int blocks_per_seq_len = (seq_len + seq_len_per_block - 1) / seq_len_per_block; + dim3 blocks(blocks_per_seq_len, channels, batch_size); + dim3 threads(threads_per_block, 1, 1); + + anti_alias_activation_forward + <<>>(dst, src, up_ftr, down_ftr, alpha, beta, batch_size, channels, seq_len); + } + } +} + +extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta) +{ + // Input is a 3d tensor with dimensions [batches, channels, seq_len] + const int batches = input.size(0); + const int channels = input.size(1); + const int seq_len = input.size(2); + + // Output + auto act_options = input.options().requires_grad(false); + + torch::Tensor anti_alias_activation_results = + torch::empty({batches, channels, seq_len}, act_options); + + void *input_ptr = static_cast(input.data_ptr()); + void *up_filter_ptr = static_cast(up_filter.data_ptr()); + void *down_filter_ptr = static_cast(down_filter.data_ptr()); + void *alpha_ptr = static_cast(alpha.data_ptr()); + void *beta_ptr = static_cast(beta.data_ptr()); + void *anti_alias_activation_results_ptr = static_cast(anti_alias_activation_results.data_ptr()); + + DISPATCH_FLOAT_HALF_AND_BFLOAT( + input.scalar_type(), + "dispatch anti alias activation_forward", + dispatch_anti_alias_activation_forward( + reinterpret_cast(anti_alias_activation_results_ptr), + reinterpret_cast(input_ptr), + reinterpret_cast(up_filter_ptr), + reinterpret_cast(down_filter_ptr), + reinterpret_cast(alpha_ptr), + reinterpret_cast(beta_ptr), + batches, + channels, + seq_len);); + return anti_alias_activation_results; +} \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/compat.h b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/compat.h new file mode 100644 index 0000000000000000000000000000000000000000..25818b2edf4cb0dc9130e62c7c4de8d16a01baa5 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/compat.h @@ -0,0 +1,29 @@ +/* coding=utf-8 + * Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/*This code is copied fron NVIDIA apex: + * https://github.com/NVIDIA/apex + * with minor changes. */ + +#ifndef TORCH_CHECK +#define TORCH_CHECK AT_CHECK +#endif + +#ifdef VERSION_GE_1_3 +#define DATA_PTR data_ptr +#else +#define DATA_PTR data +#endif diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/load.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/load.py new file mode 100644 index 0000000000000000000000000000000000000000..ca5d01de398249e75e9e2298958764acb436edba --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/load.py @@ -0,0 +1,86 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import os +import pathlib +import subprocess + +from torch.utils import cpp_extension + +""" +Setting this param to a list has a problem of generating different compilation commands (with diferent order of architectures) and leading to recompilation of fused kernels. +Set it to empty stringo avoid recompilation and assign arch flags explicity in extra_cuda_cflags below +""" +os.environ["TORCH_CUDA_ARCH_LIST"] = "" + + +def load(): + # Check if cuda 11 is installed for compute capability 8.0 + cc_flag = [] + _, bare_metal_major, _ = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME) + if int(bare_metal_major) >= 11: + cc_flag.append("-gencode") + cc_flag.append("arch=compute_80,code=sm_80") + + # Build path + srcpath = pathlib.Path(__file__).parent.absolute() + buildpath = srcpath / "build" + _create_build_dir(buildpath) + + # Helper function to build the kernels. + def _cpp_extention_load_helper(name, sources, extra_cuda_flags): + return cpp_extension.load( + name=name, + sources=sources, + build_directory=buildpath, + extra_cflags=[ + "-O3", + ], + extra_cuda_cflags=[ + "-O3", + "-gencode", + "arch=compute_70,code=sm_70", + "--use_fast_math", + ] + + extra_cuda_flags + + cc_flag, + verbose=True, + ) + + extra_cuda_flags = [ + "-U__CUDA_NO_HALF_OPERATORS__", + "-U__CUDA_NO_HALF_CONVERSIONS__", + "--expt-relaxed-constexpr", + "--expt-extended-lambda", + ] + + sources = [ + srcpath / "anti_alias_activation.cpp", + srcpath / "anti_alias_activation_cuda.cu", + ] + anti_alias_activation_cuda = _cpp_extention_load_helper( + "anti_alias_activation_cuda", sources, extra_cuda_flags + ) + + return anti_alias_activation_cuda + + +def _get_cuda_bare_metal_version(cuda_dir): + raw_output = subprocess.check_output( + [cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True + ) + output = raw_output.split() + release_idx = output.index("release") + 1 + release = output[release_idx].split(".") + bare_metal_major = release[0] + bare_metal_minor = release[1][0] + + return raw_output, bare_metal_major, bare_metal_minor + + +def _create_build_dir(buildpath): + try: + os.mkdir(buildpath) + except OSError: + if not os.path.isdir(buildpath): + print(f"Creation of the build directory {buildpath} failed") diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/type_shim.h b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/type_shim.h new file mode 100644 index 0000000000000000000000000000000000000000..5db7e8a397e982d4d30d16ab6060814b98b7ab83 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/type_shim.h @@ -0,0 +1,92 @@ +/* coding=utf-8 + * Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include +#include "compat.h" + +#define DISPATCH_FLOAT_HALF_AND_BFLOAT(TYPE, NAME, ...) \ + switch (TYPE) \ + { \ + case at::ScalarType::Float: \ + { \ + using scalar_t = float; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::Half: \ + { \ + using scalar_t = at::Half; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::BFloat16: \ + { \ + using scalar_t = at::BFloat16; \ + __VA_ARGS__; \ + break; \ + } \ + default: \ + AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'"); \ + } + +#define DISPATCH_FLOAT_HALF_AND_BFLOAT_INOUT_TYPES(TYPEIN, TYPEOUT, NAME, ...) \ + switch (TYPEIN) \ + { \ + case at::ScalarType::Float: \ + { \ + using scalar_t_in = float; \ + switch (TYPEOUT) \ + { \ + case at::ScalarType::Float: \ + { \ + using scalar_t_out = float; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::Half: \ + { \ + using scalar_t_out = at::Half; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::BFloat16: \ + { \ + using scalar_t_out = at::BFloat16; \ + __VA_ARGS__; \ + break; \ + } \ + default: \ + AT_ERROR(#NAME, " not implemented for '", toString(TYPEOUT), "'"); \ + } \ + break; \ + } \ + case at::ScalarType::Half: \ + { \ + using scalar_t_in = at::Half; \ + using scalar_t_out = at::Half; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::BFloat16: \ + { \ + using scalar_t_in = at::BFloat16; \ + using scalar_t_out = at::BFloat16; \ + __VA_ARGS__; \ + break; \ + } \ + default: \ + AT_ERROR(#NAME, " not implemented for '", toString(TYPEIN), "'"); \ + } diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8f756ed83f87f9839e457b240f60469bc187707d --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/__init__.py @@ -0,0 +1,6 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +from .filter import * +from .resample import * +from .act import * diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/act.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/act.py new file mode 100644 index 0000000000000000000000000000000000000000..1dfb03b1ee0bf1d4690ddad79ce5c567f173b198 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/act.py @@ -0,0 +1,32 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch.nn as nn + +from .....ext.bigvgan_v2.alias_free_activation.torch.resample import (DownSample1d, UpSample1d) + + +class Activation1d(nn.Module): + + def __init__( + self, + activation, + up_ratio: int = 2, + down_ratio: int = 2, + up_kernel_size: int = 12, + down_kernel_size: int = 12, + ): + super().__init__() + self.up_ratio = up_ratio + self.down_ratio = down_ratio + self.act = activation + self.upsample = UpSample1d(up_ratio, up_kernel_size) + self.downsample = DownSample1d(down_ratio, down_kernel_size) + + # x: [B,C,T] + def forward(self, x): + x = self.upsample(x) + x = self.act(x) + x = self.downsample(x) + + return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/filter.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/filter.py new file mode 100644 index 0000000000000000000000000000000000000000..0fa35b0d5ddf8d6cb04cd9d47364ca033cebcd32 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/filter.py @@ -0,0 +1,101 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch +import torch.nn as nn +import torch.nn.functional as F +import math + +if "sinc" in dir(torch): + sinc = torch.sinc +else: + # This code is adopted from adefossez's julius.core.sinc under the MIT License + # https://adefossez.github.io/julius/julius/core.html + # LICENSE is in incl_licenses directory. + def sinc(x: torch.Tensor): + """ + Implementation of sinc, i.e. sin(pi * x) / (pi * x) + __Warning__: Different to julius.sinc, the input is multiplied by `pi`! + """ + return torch.where( + x == 0, + torch.tensor(1.0, device=x.device, dtype=x.dtype), + torch.sin(math.pi * x) / math.pi / x, + ) + + +# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License +# https://adefossez.github.io/julius/julius/lowpass.html +# LICENSE is in incl_licenses directory. +def kaiser_sinc_filter1d( + cutoff, half_width, kernel_size +): # return filter [1,1,kernel_size] + even = kernel_size % 2 == 0 + half_size = kernel_size // 2 + + # For kaiser window + delta_f = 4 * half_width + A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95 + if A > 50.0: + beta = 0.1102 * (A - 8.7) + elif A >= 21.0: + beta = 0.5842 * (A - 21) ** 0.4 + 0.07886 * (A - 21.0) + else: + beta = 0.0 + window = torch.kaiser_window(kernel_size, beta=beta, periodic=False) + + # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio + if even: + time = torch.arange(-half_size, half_size) + 0.5 + else: + time = torch.arange(kernel_size) - half_size + if cutoff == 0: + filter_ = torch.zeros_like(time) + else: + filter_ = 2 * cutoff * window * sinc(2 * cutoff * time) + """ + Normalize filter to have sum = 1, otherwise we will have a small leakage of the constant component in the input signal. + """ + filter_ /= filter_.sum() + filter = filter_.view(1, 1, kernel_size) + + return filter + + +class LowPassFilter1d(nn.Module): + def __init__( + self, + cutoff=0.5, + half_width=0.6, + stride: int = 1, + padding: bool = True, + padding_mode: str = "replicate", + kernel_size: int = 12, + ): + """ + kernel_size should be even number for stylegan3 setup, in this implementation, odd number is also possible. + """ + super().__init__() + if cutoff < -0.0: + raise ValueError("Minimum cutoff must be larger than zero.") + if cutoff > 0.5: + raise ValueError("A cutoff above 0.5 does not make sense.") + self.kernel_size = kernel_size + self.even = kernel_size % 2 == 0 + self.pad_left = kernel_size // 2 - int(self.even) + self.pad_right = kernel_size // 2 + self.stride = stride + self.padding = padding + self.padding_mode = padding_mode + filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size) + self.register_buffer("filter", filter) + + # Input [B, C, T] + def forward(self, x): + _, C, _ = x.shape + + if self.padding: + x = F.pad(x, (self.pad_left, self.pad_right), mode=self.padding_mode) + out = F.conv1d(x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C) + + return out diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/resample.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/resample.py new file mode 100644 index 0000000000000000000000000000000000000000..2b7ca1002f427d3685f25e14f55e399ca7246e5b --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/alias_free_activation/torch/resample.py @@ -0,0 +1,54 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch.nn as nn +from torch.nn import functional as F + +from .....ext.bigvgan_v2.alias_free_activation.torch.filter import (LowPassFilter1d, + kaiser_sinc_filter1d) + + +class UpSample1d(nn.Module): + + def __init__(self, ratio=2, kernel_size=None): + super().__init__() + self.ratio = ratio + self.kernel_size = (int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size) + self.stride = ratio + self.pad = self.kernel_size // ratio - 1 + self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2 + self.pad_right = (self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2) + filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio, + half_width=0.6 / ratio, + kernel_size=self.kernel_size) + self.register_buffer("filter", filter) + + # x: [B, C, T] + def forward(self, x): + _, C, _ = x.shape + + x = F.pad(x, (self.pad, self.pad), mode="replicate") + x = self.ratio * F.conv_transpose1d( + x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C) + x = x[..., self.pad_left:-self.pad_right] + + return x + + +class DownSample1d(nn.Module): + + def __init__(self, ratio=2, kernel_size=None): + super().__init__() + self.ratio = ratio + self.kernel_size = (int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size) + self.lowpass = LowPassFilter1d( + cutoff=0.5 / ratio, + half_width=0.6 / ratio, + stride=ratio, + kernel_size=self.kernel_size, + ) + + def forward(self, x): + xx = self.lowpass(x) + + return xx diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/bigvgan.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/bigvgan.py new file mode 100644 index 0000000000000000000000000000000000000000..b993f1b2df4d504607246285f94fb1a0bfe8d71f --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/bigvgan.py @@ -0,0 +1,444 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import json +import os +from pathlib import Path +from typing import Dict, Optional, Union + +import torch +import torch.nn as nn +from huggingface_hub import PyTorchModelHubMixin, hf_hub_download +from torch.nn import Conv1d, ConvTranspose1d +from torch.nn.utils.parametrizations import weight_norm +from torch.nn.utils.parametrize import remove_parametrizations + +from ...ext.bigvgan_v2 import activations +from ...ext.bigvgan_v2.alias_free_activation.torch.act import \ + Activation1d as TorchActivation1d +from ...ext.bigvgan_v2.env import AttrDict +from ...ext.bigvgan_v2.utils import get_padding, init_weights + + +def load_hparams_from_json(path) -> AttrDict: + with open(path) as f: + data = f.read() + return AttrDict(json.loads(data)) + + +class AMPBlock1(torch.nn.Module): + """ + AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer. + AMPBlock1 has additional self.convs2 that contains additional Conv1d layers with a fixed dilation=1 followed by each layer in self.convs1 + + Args: + h (AttrDict): Hyperparameters. + channels (int): Number of convolution channels. + kernel_size (int): Size of the convolution kernel. Default is 3. + dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5). + activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None. + """ + + def __init__( + self, + h: AttrDict, + channels: int, + kernel_size: int = 3, + dilation: tuple = (1, 3, 5), + activation: str = None, + ): + super().__init__() + + self.h = h + + self.convs1 = nn.ModuleList([ + weight_norm( + Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=d, + padding=get_padding(kernel_size, d), + )) for d in dilation + ]) + self.convs1.apply(init_weights) + + self.convs2 = nn.ModuleList([ + weight_norm( + Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=1, + padding=get_padding(kernel_size, 1), + )) for _ in range(len(dilation)) + ]) + self.convs2.apply(init_weights) + + self.num_layers = len(self.convs1) + len(self.convs2) # Total number of conv layers + + # Select which Activation1d, lazy-load cuda version to ensure backward compatibility + if self.h.get("use_cuda_kernel", False): + from alias_free_activation.cuda.activation1d import \ + Activation1d as CudaActivation1d + + Activation1d = CudaActivation1d + else: + Activation1d = TorchActivation1d + + # Activation functions + if activation == "snake": + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.Snake(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + elif activation == "snakebeta": + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + def forward(self, x): + acts1, acts2 = self.activations[::2], self.activations[1::2] + for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2): + xt = a1(x) + xt = c1(xt) + xt = a2(xt) + xt = c2(xt) + x = xt + x + + return x + + def remove_weight_norm(self): + for l in self.convs1: + remove_parametrizations(l, 'weight') + for l in self.convs2: + remove_parametrizations(l, 'weight') + + +class AMPBlock2(torch.nn.Module): + """ + AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer. + Unlike AMPBlock1, AMPBlock2 does not contain extra Conv1d layers with fixed dilation=1 + + Args: + h (AttrDict): Hyperparameters. + channels (int): Number of convolution channels. + kernel_size (int): Size of the convolution kernel. Default is 3. + dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5). + activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None. + """ + + def __init__( + self, + h: AttrDict, + channels: int, + kernel_size: int = 3, + dilation: tuple = (1, 3, 5), + activation: str = None, + ): + super().__init__() + + self.h = h + + self.convs = nn.ModuleList([ + weight_norm( + Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=d, + padding=get_padding(kernel_size, d), + )) for d in dilation + ]) + self.convs.apply(init_weights) + + self.num_layers = len(self.convs) # Total number of conv layers + + # Select which Activation1d, lazy-load cuda version to ensure backward compatibility + if self.h.get("use_cuda_kernel", False): + from alias_free_activation.cuda.activation1d import \ + Activation1d as CudaActivation1d + + Activation1d = CudaActivation1d + else: + Activation1d = TorchActivation1d + + # Activation functions + if activation == "snake": + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.Snake(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + elif activation == "snakebeta": + self.activations = nn.ModuleList([ + Activation1d( + activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale)) + for _ in range(self.num_layers) + ]) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + def forward(self, x): + for c, a in zip(self.convs, self.activations): + xt = a(x) + xt = c(xt) + x = xt + x + return x + + def remove_weight_norm(self): + for l in self.convs: + remove_weight_norm(l) + + +class BigVGAN( + torch.nn.Module, + PyTorchModelHubMixin, + library_name="bigvgan", + repo_url="https://github.com/NVIDIA/BigVGAN", + docs_url="https://github.com/NVIDIA/BigVGAN/blob/main/README.md", + pipeline_tag="audio-to-audio", + license="mit", + tags=["neural-vocoder", "audio-generation", "arxiv:2206.04658"], +): + """ + BigVGAN is a neural vocoder model that applies anti-aliased periodic activation for residual blocks (resblocks). + New in BigVGAN-v2: it can optionally use optimized CUDA kernels for AMP (anti-aliased multi-periodicity) blocks. + + Args: + h (AttrDict): Hyperparameters. + use_cuda_kernel (bool): If set to True, loads optimized CUDA kernels for AMP. This should be used for inference only, as training is not supported with CUDA kernels. + + Note: + - The `use_cuda_kernel` parameter should be used for inference only, as training with CUDA kernels is not supported. + - Ensure that the activation function is correctly specified in the hyperparameters (h.activation). + """ + + def __init__(self, h: AttrDict, use_cuda_kernel: bool = False): + super().__init__() + self.h = h + self.h["use_cuda_kernel"] = use_cuda_kernel + + # Select which Activation1d, lazy-load cuda version to ensure backward compatibility + if self.h.get("use_cuda_kernel", False): + from alias_free_activation.cuda.activation1d import \ + Activation1d as CudaActivation1d + + Activation1d = CudaActivation1d + else: + Activation1d = TorchActivation1d + + self.num_kernels = len(h.resblock_kernel_sizes) + self.num_upsamples = len(h.upsample_rates) + + # Pre-conv + self.conv_pre = weight_norm(Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)) + + # Define which AMPBlock to use. BigVGAN uses AMPBlock1 as default + if h.resblock == "1": + resblock_class = AMPBlock1 + elif h.resblock == "2": + resblock_class = AMPBlock2 + else: + raise ValueError( + f"Incorrect resblock class specified in hyperparameters. Got {h.resblock}") + + # Transposed conv-based upsamplers. does not apply anti-aliasing + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)): + self.ups.append( + nn.ModuleList([ + weight_norm( + ConvTranspose1d( + h.upsample_initial_channel // (2**i), + h.upsample_initial_channel // (2**(i + 1)), + k, + u, + padding=(k - u) // 2, + )) + ])) + + # Residual blocks using anti-aliased multi-periodicity composition modules (AMP) + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = h.upsample_initial_channel // (2**(i + 1)) + for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)): + self.resblocks.append(resblock_class(h, ch, k, d, activation=h.activation)) + + # Post-conv + activation_post = (activations.Snake(ch, alpha_logscale=h.snake_logscale) + if h.activation == "snake" else + (activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale) + if h.activation == "snakebeta" else None)) + if activation_post is None: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + self.activation_post = Activation1d(activation=activation_post) + + # Whether to use bias for the final conv_post. Default to True for backward compatibility + self.use_bias_at_final = h.get("use_bias_at_final", True) + self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3, bias=self.use_bias_at_final)) + + # Weight initialization + for i in range(len(self.ups)): + self.ups[i].apply(init_weights) + self.conv_post.apply(init_weights) + + # Final tanh activation. Defaults to True for backward compatibility + self.use_tanh_at_final = h.get("use_tanh_at_final", True) + + def forward(self, x): + # Pre-conv + x = self.conv_pre(x) + + for i in range(self.num_upsamples): + # Upsampling + for i_up in range(len(self.ups[i])): + x = self.ups[i][i_up](x) + # AMP blocks + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + + # Post-conv + x = self.activation_post(x) + x = self.conv_post(x) + # Final tanh activation + if self.use_tanh_at_final: + x = torch.tanh(x) + else: + x = torch.clamp(x, min=-1.0, max=1.0) # Bound the output to [-1, 1] + + return x + + def remove_weight_norm(self): + try: + print("Removing weight norm...") + for l in self.ups: + for l_i in l: + remove_parametrizations(l_i, 'weight') + for l in self.resblocks: + l.remove_weight_norm() + remove_parametrizations(self.conv_pre, 'weight') + remove_parametrizations(self.conv_post, 'weight') + except ValueError: + print("[INFO] Model already removed weight norm. Skipping!") + pass + + # Additional methods for huggingface_hub support + def _save_pretrained(self, save_directory: Path) -> None: + """Save weights and config.json from a Pytorch model to a local directory.""" + + model_path = save_directory / "bigvgan_generator.pt" + torch.save({"generator": self.state_dict()}, model_path) + + config_path = save_directory / "config.json" + with open(config_path, "w") as config_file: + json.dump(self.h, config_file, indent=4) + + @classmethod + def _from_pretrained( + cls, + *, + model_id: str, + revision: str, + cache_dir: str, + force_download: bool, + proxies: Optional[Dict] = None, + resume_download: bool = False, + local_files_only: bool, + token: Union[str, bool, None], + map_location: str = "cpu", # Additional argument + strict: bool = False, # Additional argument + use_cuda_kernel: bool = False, + **model_kwargs, + ): + """Load Pytorch pretrained weights and return the loaded model.""" + + # Build kwargs for hf_hub_download, excluding params removed in huggingface_hub 1.0 + hf_hub_kwargs = dict( + revision=revision, + cache_dir=cache_dir, + force_download=force_download, + token=token, + local_files_only=local_files_only, + ) + # proxies and resume_download were removed in huggingface_hub 1.0 + # Only include them for backwards compatibility with older versions + import inspect + hf_download_params = inspect.signature(hf_hub_download).parameters + if "proxies" in hf_download_params: + hf_hub_kwargs["proxies"] = proxies + if "resume_download" in hf_download_params: + hf_hub_kwargs["resume_download"] = resume_download + + # Download and load hyperparameters (h) used by BigVGAN + if os.path.isdir(model_id): + print("Loading config.json from local directory") + config_file = os.path.join(model_id, "config.json") + else: + config_file = hf_hub_download( + repo_id=model_id, + filename="config.json", + **hf_hub_kwargs, + ) + h = load_hparams_from_json(config_file) + + # instantiate BigVGAN using h + if use_cuda_kernel: + print( + f"[WARNING] You have specified use_cuda_kernel=True during BigVGAN.from_pretrained(). Only inference is supported (training is not implemented)!" + ) + print( + f"[WARNING] You need nvcc and ninja installed in your system that matches your PyTorch build is using to build the kernel. If not, the model will fail to initialize or generate incorrect waveform!" + ) + print( + f"[WARNING] For detail, see the official GitHub repository: https://github.com/NVIDIA/BigVGAN?tab=readme-ov-file#using-custom-cuda-kernel-for-synthesis" + ) + model = cls(h, use_cuda_kernel=use_cuda_kernel) + + # Download and load pretrained generator weight + if os.path.isdir(model_id): + print("Loading weights from local directory") + model_file = os.path.join(model_id, "bigvgan_generator.pt") + else: + print(f"Loading weights from {model_id}") + model_file = hf_hub_download( + repo_id=model_id, + filename="bigvgan_generator.pt", + **hf_hub_kwargs, + ) + + checkpoint_dict = torch.load(model_file, map_location=map_location, weights_only=True) + + try: + model.load_state_dict(checkpoint_dict["generator"]) + except RuntimeError: + print( + f"[INFO] the pretrained checkpoint does not contain weight norm. Loading the checkpoint after removing weight norm!" + ) + model.remove_weight_norm() + model.load_state_dict(checkpoint_dict["generator"]) + + return model diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/env.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/env.py new file mode 100644 index 0000000000000000000000000000000000000000..b8be238d4db710c8c9a338d336baea0138f18d1f --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/env.py @@ -0,0 +1,18 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import os +import shutil + + +class AttrDict(dict): + def __init__(self, *args, **kwargs): + super(AttrDict, self).__init__(*args, **kwargs) + self.__dict__ = self + + +def build_env(config, config_name, path): + t_path = os.path.join(path, config_name) + if config != t_path: + os.makedirs(path, exist_ok=True) + shutil.copyfile(config, os.path.join(path, config_name)) \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_1 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_1 new file mode 100644 index 0000000000000000000000000000000000000000..5afae394d6b37da0e12ba6b290d2512687f421ac --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_1 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Jungil Kong + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_2 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_2 new file mode 100644 index 0000000000000000000000000000000000000000..322b758863c4219be68291ae3826218baa93cb4c --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_2 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Edward Dixon + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_3 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_3 new file mode 100644 index 0000000000000000000000000000000000000000..56ee3c8c4cc2b4b32e0975d17258f9ba515fdbcc --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_3 @@ -0,0 +1,201 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_4 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_4 new file mode 100644 index 0000000000000000000000000000000000000000..48fd1a1ba8d81a94b6c7d1c2ff1a1f307cc5371d --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_4 @@ -0,0 +1,29 @@ +BSD 3-Clause License + +Copyright (c) 2019, Seungwon Park 박승원 +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +3. Neither the name of the copyright holder nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_5 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_5 new file mode 100644 index 0000000000000000000000000000000000000000..01ae5538e6b7c787bb4f5d6f2cd9903520d6e465 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_5 @@ -0,0 +1,16 @@ +Copyright 2020 Alexandre Défossez + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and +associated documentation files (the "Software"), to deal in the Software without restriction, +including without limitation the rights to use, copy, modify, merge, publish, distribute, +sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or +substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT +NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND +NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, +DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_6 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_6 new file mode 100644 index 0000000000000000000000000000000000000000..2569ec0b6c85f94f3cd071ba16e9028ccf156be2 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_6 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023-present, Descript + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_7 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_7 new file mode 100644 index 0000000000000000000000000000000000000000..c37bdaf99c6921f5849425d546069e972f52d7fa --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_7 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Charactr Inc. + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_8 b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_8 new file mode 100644 index 0000000000000000000000000000000000000000..ab3d7ffe795779f54e339078e4e752ad9019aae8 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/incl_licenses/LICENSE_8 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Amphion + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..3b1d41670fa1ee257b2ed22c61086ba7a32c7cb0 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/bigvgan_v2/utils.py @@ -0,0 +1,31 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import os + +import torch +from torch.nn.utils import weight_norm + + +def init_weights(m, mean=0.0, std=0.01): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + m.weight.data.normal_(mean, std) + + +def apply_weight_norm(m): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + weight_norm(m) + + +def get_padding(kernel_size, dilation=1): + return int((kernel_size * dilation - dilation) / 2) + + +def load_checkpoint(filepath, device): + assert os.path.isfile(filepath) + print(f"Loading '{filepath}'") + checkpoint_dict = torch.load(filepath, map_location=device) + print("Complete.") + return checkpoint_dict diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/mel_converter.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/mel_converter.py new file mode 100644 index 0000000000000000000000000000000000000000..6fc589c9468e077fc580965db250fd502e229672 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/mel_converter.py @@ -0,0 +1,82 @@ +# Reference: # https://github.com/bytedance/Make-An-Audio-2 + +import torch +import torch.nn as nn +from librosa.filters import mel as librosa_mel_fn + + +def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, norm_fn=torch.log10): + return norm_fn(torch.clamp(x, min=clip_val) * C) + + +def spectral_normalize_torch(magnitudes, norm_fn): + output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn) + return output + + +class MelConverter(nn.Module): + + def __init__( + self, + *, + sampling_rate: float = 16_000, + n_fft: int = 1024, + num_mels: int = 80, + hop_size: int = 256, + win_size: int = 1024, + fmin: float = 0, + fmax: float = 8_000, + norm_fn=torch.log10, + ): + super().__init__() + self.sampling_rate = sampling_rate + self.n_fft = n_fft + self.num_mels = num_mels + self.hop_size = hop_size + self.win_size = win_size + self.fmin = fmin + self.fmax = fmax + self.norm_fn = norm_fn + + mel = librosa_mel_fn(sr=self.sampling_rate, + n_fft=self.n_fft, + n_mels=self.num_mels, + fmin=self.fmin, + fmax=self.fmax) + mel_basis = torch.from_numpy(mel).float() + hann_window = torch.hann_window(self.win_size) + + self.register_buffer('mel_basis', mel_basis) + self.register_buffer('hann_window', hann_window) + + @property + def device(self): + return self.mel_basis.device + + def forward(self, waveform: torch.Tensor, center: bool = False) -> torch.Tensor: + waveform = waveform.clamp(min=-1., max=1.).to(self.device) + + waveform = torch.nn.functional.pad( + waveform.unsqueeze(1), + [int((self.n_fft - self.hop_size) / 2), + int((self.n_fft - self.hop_size) / 2)], + mode='reflect') + waveform = waveform.squeeze(1) + + spec = torch.stft(waveform, + self.n_fft, + hop_length=self.hop_size, + win_length=self.win_size, + window=self.hann_window, + center=center, + pad_mode='reflect', + normalized=False, + onesided=True, + return_complex=True) + + spec = torch.view_as_real(spec) + spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9)) + spec = torch.matmul(self.mel_basis, spec) + spec = spectral_normalize_torch(spec, self.norm_fn) + + return spec diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/rotary_embeddings.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/rotary_embeddings.py new file mode 100644 index 0000000000000000000000000000000000000000..1ea9d56278cb68b7577ed13148227c30ed98fd02 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/rotary_embeddings.py @@ -0,0 +1,35 @@ +from typing import Union + +import torch +from einops import rearrange +from torch import Tensor + +# Ref: https://github.com/black-forest-labs/flux/blob/main/src/flux/math.py +# Ref: https://github.com/lucidrains/rotary-embedding-torch + + +def compute_rope_rotations(length: int, + dim: int, + theta: int, + *, + freq_scaling: float = 1.0, + device: Union[torch.device, str] = 'cpu') -> Tensor: + assert dim % 2 == 0 + + with torch.amp.autocast(device_type='cuda', enabled=False): + pos = torch.arange(length, dtype=torch.float32, device=device) + freqs = 1.0 / (theta**(torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim)) + freqs *= freq_scaling + + rot = torch.einsum('..., f -> ... f', pos, freqs) + rot = torch.stack([torch.cos(rot), -torch.sin(rot), torch.sin(rot), torch.cos(rot)], dim=-1) + rot = rearrange(rot, 'n d (i j) -> 1 n d i j', i=2, j=2) + return rot + + +def apply_rope(x: Tensor, rot: Tensor) -> tuple[Tensor, Tensor]: + with torch.amp.autocast(device_type='cuda', enabled=False): + _x = x.float() + _x = _x.view(*_x.shape[:-1], -1, 1, 2) + x_out = rot[..., 0] * _x[..., 0] + rot[..., 1] * _x[..., 1] + return x_out.reshape(*x.shape).to(dtype=x.dtype) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/stft_converter.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/stft_converter.py new file mode 100644 index 0000000000000000000000000000000000000000..62922067ef3b1d3b8727ec39e7d664ccb304d9fe --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/stft_converter.py @@ -0,0 +1,183 @@ +# Reference: # https://github.com/bytedance/Make-An-Audio-2 + +import torch +import torch.nn as nn +import torchaudio +from einops import rearrange +from librosa.filters import mel as librosa_mel_fn + + +def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, norm_fn=torch.log10): + return norm_fn(torch.clamp(x, min=clip_val) * C) + + +def spectral_normalize_torch(magnitudes, norm_fn): + output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn) + return output + + +class STFTConverter(nn.Module): + + def __init__( + self, + *, + sampling_rate: float = 16_000, + n_fft: int = 1024, + num_mels: int = 128, + hop_size: int = 256, + win_size: int = 1024, + fmin: float = 0, + fmax: float = 8_000, + norm_fn=torch.log, + ): + super().__init__() + self.sampling_rate = sampling_rate + self.n_fft = n_fft + self.num_mels = num_mels + self.hop_size = hop_size + self.win_size = win_size + self.fmin = fmin + self.fmax = fmax + self.norm_fn = norm_fn + + mel = librosa_mel_fn(sr=self.sampling_rate, + n_fft=self.n_fft, + n_mels=self.num_mels, + fmin=self.fmin, + fmax=self.fmax) + mel_basis = torch.from_numpy(mel).float() + hann_window = torch.hann_window(self.win_size) + + self.register_buffer('mel_basis', mel_basis) + self.register_buffer('hann_window', hann_window) + + @property + def device(self): + return self.hann_window.device + + def forward(self, waveform: torch.Tensor) -> torch.Tensor: + # input: batch_size * length + bs = waveform.shape[0] + waveform = waveform.clamp(min=-1., max=1.) + + spec = torch.stft(waveform, + self.n_fft, + hop_length=self.hop_size, + win_length=self.win_size, + window=self.hann_window, + center=True, + pad_mode='reflect', + normalized=False, + onesided=True, + return_complex=True) + + spec = torch.view_as_real(spec) + # print('After stft', spec.shape, spec.min(), spec.max(), spec.mean()) + + power = spec.pow(2).sum(-1) + angle = torch.atan2(spec[..., 1], spec[..., 0]) + + print('power', power.shape, power.min(), power.max(), power.mean()) + print('angle', angle.shape, angle.min(), angle.max(), angle.mean()) + + # print('mel', self.mel_basis.shape, self.mel_basis.min(), self.mel_basis.max(), + # self.mel_basis.mean()) + + # spec = rearrange(spec, 'b f t c -> (b c) f t') + + # spec = self.mel_transform(spec) + + # spec = torch.matmul(self.mel_basis, spec) + + # print('After mel', spec.shape, spec.min(), spec.max(), spec.mean()) + + # spec = spectral_normalize_torch(spec, self.norm_fn) + + # print('After norm', spec.shape, spec.min(), spec.max(), spec.mean()) + + # compute magnitude + # magnitude = torch.sqrt((spec**2).sum(-1)) + # normalize by magnitude + # scaled_magnitude = torch.log10(magnitude.clamp(min=1e-5)) * 10 + # spec = spec / magnitude.unsqueeze(-1) * scaled_magnitude.unsqueeze(-1) + + # power = torch.log10(power.clamp(min=1e-5)) * 10 + power = torch.log10(power.clamp(min=1e-5)) + + print('After scaling', power.shape, power.min(), power.max(), power.mean()) + + spec = torch.stack([power, angle], dim=-1) + + # spec = rearrange(spec, '(b c) f t -> b c f t', b=bs) + spec = rearrange(spec, 'b f t c -> b c f t', b=bs) + + # spec[:, :, 400:] = 0 + + return spec + + def invert(self, spec: torch.Tensor, length: int) -> torch.Tensor: + bs = spec.shape[0] + + # spec = rearrange(spec, 'b c f t -> (b c) f t') + # print(spec.shape, self.mel_basis.shape) + # spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution + # spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec + + # spec = self.invmel_transform(spec) + + spec = rearrange(spec, 'b c f t -> b f t c', b=bs).contiguous() + + # spec[..., 0] = 10**(spec[..., 0] / 10) + + power = spec[..., 0] + power = 10**power + + # print('After unscaling', spec[..., 0].shape, spec[..., 0].min(), spec[..., 0].max(), + # spec[..., 0].mean()) + + unit_vector = torch.stack([ + torch.cos(spec[..., 1]), + torch.sin(spec[..., 1]), + ], dim=-1) + + spec = torch.sqrt(power) * unit_vector + + # spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous() + spec = torch.view_as_complex(spec) + + waveform = torch.istft( + spec, + self.n_fft, + length=length, + hop_length=self.hop_size, + win_length=self.win_size, + window=self.hann_window, + center=True, + normalized=False, + onesided=True, + return_complex=False, + ) + + return waveform + + +if __name__ == '__main__': + + converter = STFTConverter(sampling_rate=16000) + + signal = torchaudio.load('./output/ZZ6GRocWW38_000090.wav')[0] + # resample signal at 44100 Hz + # signal = torchaudio.transforms.Resample(16_000, 44_100)(signal) + + L = signal.shape[1] + print('Input signal', signal.shape) + spec = converter(signal) + + print('Final spec', spec.shape) + + signal_recon = converter.invert(spec, length=L) + print('Output signal', signal_recon.shape, signal_recon.min(), signal_recon.max(), + signal_recon.mean()) + + print('MSE', torch.nn.functional.mse_loss(signal, signal_recon)) + torchaudio.save('./output/ZZ6GRocWW38_000090_recon.wav', signal_recon, 16000) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/stft_converter_mel.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/stft_converter_mel.py new file mode 100644 index 0000000000000000000000000000000000000000..f6b32d4cb9a23cd74f723e7d8307fd82fa1abba0 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/stft_converter_mel.py @@ -0,0 +1,234 @@ +# Reference: # https://github.com/bytedance/Make-An-Audio-2 + +import torch +import torch.nn as nn +import torchaudio +from einops import rearrange +from librosa.filters import mel as librosa_mel_fn + + +def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, norm_fn=torch.log10): + return norm_fn(torch.clamp(x, min=clip_val) * C) + + +def spectral_normalize_torch(magnitudes, norm_fn): + output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn) + return output + + +class STFTConverter(nn.Module): + + def __init__( + self, + *, + sampling_rate: float = 16_000, + n_fft: int = 1024, + num_mels: int = 128, + hop_size: int = 256, + win_size: int = 1024, + fmin: float = 0, + fmax: float = 8_000, + norm_fn=torch.log, + ): + super().__init__() + self.sampling_rate = sampling_rate + self.n_fft = n_fft + self.num_mels = num_mels + self.hop_size = hop_size + self.win_size = win_size + self.fmin = fmin + self.fmax = fmax + self.norm_fn = norm_fn + + mel = librosa_mel_fn(sr=self.sampling_rate, + n_fft=self.n_fft, + n_mels=self.num_mels, + fmin=self.fmin, + fmax=self.fmax) + mel_basis = torch.from_numpy(mel).float() + hann_window = torch.hann_window(self.win_size) + + self.register_buffer('mel_basis', mel_basis) + self.register_buffer('hann_window', hann_window) + + @property + def device(self): + return self.hann_window.device + + def forward(self, waveform: torch.Tensor) -> torch.Tensor: + # input: batch_size * length + bs = waveform.shape[0] + waveform = waveform.clamp(min=-1., max=1.) + + spec = torch.stft(waveform, + self.n_fft, + hop_length=self.hop_size, + win_length=self.win_size, + window=self.hann_window, + center=True, + pad_mode='reflect', + normalized=False, + onesided=True, + return_complex=True) + + spec = torch.view_as_real(spec) + # print('After stft', spec.shape, spec.min(), spec.max(), spec.mean()) + + power = (spec.pow(2).sum(-1))**(0.5) + angle = torch.atan2(spec[..., 1], spec[..., 0]) + + print('power 1', power.shape, power.min(), power.max(), power.mean()) + print('angle 1', angle.shape, angle.min(), angle.max(), angle.mean(), angle[:, :2, :2]) + + # print('mel', self.mel_basis.shape, self.mel_basis.min(), self.mel_basis.max(), + # self.mel_basis.mean()) + + # spec = self.mel_transform(spec) + + # power = torch.matmul(self.mel_basis, power) + + spec = rearrange(spec, 'b f t c -> (b c) f t') + spec = self.mel_basis.unsqueeze(0) @ spec + spec = rearrange(spec, '(b c) f t -> b f t c', b=bs) + + power = (spec.pow(2).sum(-1))**(0.5) + angle = torch.atan2(spec[..., 1], spec[..., 0]) + + print('power', power.shape, power.min(), power.max(), power.mean()) + print('angle', angle.shape, angle.min(), angle.max(), angle.mean(), angle[:, :2, :2]) + + # print('After mel', spec.shape, spec.min(), spec.max(), spec.mean()) + + # spec = spectral_normalize_torch(spec, self.norm_fn) + + # print('After norm', spec.shape, spec.min(), spec.max(), spec.mean()) + + # compute magnitude + # magnitude = torch.sqrt((spec**2).sum(-1)) + # normalize by magnitude + # scaled_magnitude = torch.log10(magnitude.clamp(min=1e-5)) * 10 + # spec = spec / magnitude.unsqueeze(-1) * scaled_magnitude.unsqueeze(-1) + + # power = torch.log10(power.clamp(min=1e-5)) * 10 + power = torch.log10(power.clamp(min=1e-8)) + + print('After scaling', power.shape, power.min(), power.max(), power.mean()) + + # spec = torch.stack([power, angle], dim=-1) + + # spec = rearrange(spec, '(b c) f t -> b c f t', b=bs) + # spec = rearrange(spec, 'b f t c -> b c f t', b=bs) + + # spec[:, :, 400:] = 0 + + return power, angle + # return spec[..., 0], spec[..., 1] + + def invert(self, spec: torch.Tensor, length: int) -> torch.Tensor: + + power, angle = spec + + bs = power.shape[0] + + # spec = rearrange(spec, 'b c f t -> (b c) f t') + # print(spec.shape, self.mel_basis.shape) + # spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution + # spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec + + # spec = self.invmel_transform(spec) + + # spec = rearrange(spec, 'b c f t -> b f t c', b=bs).contiguous() + + # spec[..., 0] = 10**(spec[..., 0] / 10) + + # power = spec[..., 0] + power = 10**power + + # print('After unscaling', spec[..., 0].shape, spec[..., 0].min(), spec[..., 0].max(), + # spec[..., 0].mean()) + + unit_vector = torch.stack([ + torch.cos(angle), + torch.sin(angle), + ], dim=-1) + + spec = power.unsqueeze(-1) * unit_vector + + # power = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), power).solution + spec = rearrange(spec, 'b f t c -> (b c) f t') + spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec + # spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution + spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous() + + power = (spec.pow(2).sum(-1))**(0.5) + angle = torch.atan2(spec[..., 1], spec[..., 0]) + + print('power 2', power.shape, power.min(), power.max(), power.mean()) + print('angle 2', angle.shape, angle.min(), angle.max(), angle.mean(), angle[:, :2, :2]) + + # spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous() + spec = torch.view_as_complex(spec) + + waveform = torch.istft( + spec, + self.n_fft, + length=length, + hop_length=self.hop_size, + win_length=self.win_size, + window=self.hann_window, + center=True, + normalized=False, + onesided=True, + return_complex=False, + ) + + return waveform + + +if __name__ == '__main__': + + converter = STFTConverter(sampling_rate=16000) + + signal = torchaudio.load('./output/ZZ6GRocWW38_000090.wav')[0] + # resample signal at 44100 Hz + # signal = torchaudio.transforms.Resample(16_000, 44_100)(signal) + + L = signal.shape[1] + print('Input signal', signal.shape) + spec = converter(signal) + + power, angle = spec + + # print(power.shape, angle.shape) + # print(power, power.min(), power.max(), power.mean()) + # power = power.clamp(-1, 1) + # angle = angle.clamp(-1, 1) + + import matplotlib.pyplot as plt + + # Visualize power + plt.figure() + plt.imshow(power[0].detach().numpy(), aspect='auto', origin='lower') + plt.colorbar() + plt.title('Power') + plt.xlabel('Time') + plt.ylabel('Frequency') + plt.savefig('./output/power.png') + + # Visualize angle + plt.figure() + plt.imshow(angle[0].detach().numpy(), aspect='auto', origin='lower') + plt.colorbar() + plt.title('Angle') + plt.xlabel('Time') + plt.ylabel('Frequency') + plt.savefig('./output/angle.png') + + # print('Final spec', spec.shape) + + signal_recon = converter.invert(spec, length=L) + print('Output signal', signal_recon.shape, signal_recon.min(), signal_recon.max(), + signal_recon.mean()) + + print('MSE', torch.nn.functional.mse_loss(signal, signal_recon)) + torchaudio.save('./output/ZZ6GRocWW38_000090_recon.wav', signal_recon, 16000) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/LICENSE b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..2f70bf24b6f45f458998bdf5746376c4832352ea --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2024 Vladimir Iashin + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..46ae4ff58457684d00454b9a7ff533177cf68898 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/__init__.py @@ -0,0 +1 @@ +from ...ext.synchformer.synchformer import Synchformer diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/divided_224_16x4.yaml b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/divided_224_16x4.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f9d20b76302a8af7928391643bd4b2d184e970aa --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/divided_224_16x4.yaml @@ -0,0 +1,84 @@ +TRAIN: + ENABLE: True + DATASET: Ssv2 + BATCH_SIZE: 32 + EVAL_PERIOD: 5 + CHECKPOINT_PERIOD: 5 + AUTO_RESUME: True + CHECKPOINT_EPOCH_RESET: True + CHECKPOINT_FILE_PATH: /checkpoint/fmetze/neurips_sota/40944587/checkpoints/checkpoint_epoch_00035.pyth +DATA: + NUM_FRAMES: 16 + SAMPLING_RATE: 4 + TRAIN_JITTER_SCALES: [256, 320] + TRAIN_CROP_SIZE: 224 + TEST_CROP_SIZE: 224 + INPUT_CHANNEL_NUM: [3] + MEAN: [0.5, 0.5, 0.5] + STD: [0.5, 0.5, 0.5] + PATH_TO_DATA_DIR: /private/home/mandelapatrick/slowfast/data/ssv2 + PATH_PREFIX: /datasets01/SomethingV2/092720/20bn-something-something-v2-frames + INV_UNIFORM_SAMPLE: True + RANDOM_FLIP: False + REVERSE_INPUT_CHANNEL: True + USE_RAND_AUGMENT: True + RE_PROB: 0.0 + USE_REPEATED_AUG: False + USE_RANDOM_RESIZE_CROPS: False + COLORJITTER: False + GRAYSCALE: False + GAUSSIAN: False +SOLVER: + BASE_LR: 1e-4 + LR_POLICY: steps_with_relative_lrs + LRS: [1, 0.1, 0.01] + STEPS: [0, 20, 30] + MAX_EPOCH: 35 + MOMENTUM: 0.9 + WEIGHT_DECAY: 5e-2 + WARMUP_EPOCHS: 0.0 + OPTIMIZING_METHOD: adamw + USE_MIXED_PRECISION: True + SMOOTHING: 0.2 +SLOWFAST: + ALPHA: 8 +VIT: + PATCH_SIZE: 16 + PATCH_SIZE_TEMP: 2 + CHANNELS: 3 + EMBED_DIM: 768 + DEPTH: 12 + NUM_HEADS: 12 + MLP_RATIO: 4 + QKV_BIAS: True + VIDEO_INPUT: True + TEMPORAL_RESOLUTION: 8 + USE_MLP: True + DROP: 0.0 + POS_DROPOUT: 0.0 + DROP_PATH: 0.2 + IM_PRETRAINED: True + HEAD_DROPOUT: 0.0 + HEAD_ACT: tanh + PRETRAINED_WEIGHTS: vit_1k + ATTN_LAYER: divided +MODEL: + NUM_CLASSES: 174 + ARCH: slow + MODEL_NAME: VisionTransformer + LOSS_FUNC: cross_entropy +TEST: + ENABLE: True + DATASET: Ssv2 + BATCH_SIZE: 64 + NUM_ENSEMBLE_VIEWS: 1 + NUM_SPATIAL_CROPS: 3 +DATA_LOADER: + NUM_WORKERS: 4 + PIN_MEMORY: True +NUM_GPUS: 8 +NUM_SHARDS: 4 +RNG_SEED: 0 +OUTPUT_DIR: . +TENSORBOARD: + ENABLE: True diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/motionformer.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/motionformer.py new file mode 100644 index 0000000000000000000000000000000000000000..dd1f2330615a5d9f3a1773f8498bf39bd6ec02de --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/motionformer.py @@ -0,0 +1,400 @@ +import logging +from pathlib import Path + +import einops +import torch +from omegaconf import OmegaConf +from timm.layers import trunc_normal_ +from torch import nn + +from ...ext.synchformer.utils import check_if_file_exists_else_download +from ...ext.synchformer.video_model_builder import VisionTransformer + +FILE2URL = { + # cfg + 'motionformer_224_16x4.yaml': + 'https://raw.githubusercontent.com/facebookresearch/Motionformer/bf43d50/configs/SSV2/motionformer_224_16x4.yaml', + 'joint_224_16x4.yaml': + 'https://raw.githubusercontent.com/facebookresearch/Motionformer/bf43d50/configs/SSV2/joint_224_16x4.yaml', + 'divided_224_16x4.yaml': + 'https://raw.githubusercontent.com/facebookresearch/Motionformer/bf43d50/configs/SSV2/divided_224_16x4.yaml', + # ckpt + 'ssv2_motionformer_224_16x4.pyth': + 'https://dl.fbaipublicfiles.com/motionformer/ssv2_motionformer_224_16x4.pyth', + 'ssv2_joint_224_16x4.pyth': + 'https://dl.fbaipublicfiles.com/motionformer/ssv2_joint_224_16x4.pyth', + 'ssv2_divided_224_16x4.pyth': + 'https://dl.fbaipublicfiles.com/motionformer/ssv2_divided_224_16x4.pyth', +} + + +class MotionFormer(VisionTransformer): + ''' This class serves three puposes: + 1. Renames the class to MotionFormer. + 2. Downloads the cfg from the original repo and patches it if needed. + 3. Takes care of feature extraction by redefining .forward() + - if `extract_features=True` and `factorize_space_time=False`, + the output is of shape (B, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8 + - if `extract_features=True` and `factorize_space_time=True`, the output is of shape (B*S, D) + and spatial and temporal transformer encoder layers are used. + - if `extract_features=True` and `factorize_space_time=True` as well as `add_global_repr=True` + the output is of shape (B, D) and spatial and temporal transformer encoder layers + are used as well as the global representation is extracted from segments (extra pos emb + is added). + ''' + + def __init__( + self, + extract_features: bool = False, + ckpt_path: str = None, + factorize_space_time: bool = None, + agg_space_module: str = None, + agg_time_module: str = None, + add_global_repr: bool = True, + agg_segments_module: str = None, + max_segments: int = None, + ): + self.extract_features = extract_features + self.ckpt_path = ckpt_path + self.factorize_space_time = factorize_space_time + + if self.ckpt_path is not None: + check_if_file_exists_else_download(self.ckpt_path, FILE2URL) + ckpt = torch.load(self.ckpt_path, map_location='cpu') + mformer_ckpt2cfg = { + 'ssv2_motionformer_224_16x4.pyth': 'motionformer_224_16x4.yaml', + 'ssv2_joint_224_16x4.pyth': 'joint_224_16x4.yaml', + 'ssv2_divided_224_16x4.pyth': 'divided_224_16x4.yaml', + } + # init from motionformer ckpt or from our Stage I ckpt + # depending on whether the feat extractor was pre-trained on AVCLIPMoCo or not, we need to + # load the state dict differently + was_pt_on_avclip = self.ckpt_path.endswith( + '.pt') # checks if it is a stage I ckpt (FIXME: a bit generic) + if self.ckpt_path.endswith(tuple(mformer_ckpt2cfg.keys())): + cfg_fname = mformer_ckpt2cfg[Path(self.ckpt_path).name] + elif was_pt_on_avclip: + # TODO: this is a hack, we should be able to get the cfg from the ckpt (earlier ckpt didn't have it) + s1_cfg = ckpt.get('args', None) # Stage I cfg + if s1_cfg is not None: + s1_vfeat_extractor_ckpt_path = s1_cfg.model.params.vfeat_extractor.params.ckpt_path + # if the stage I ckpt was initialized from a motionformer ckpt or train from scratch + if s1_vfeat_extractor_ckpt_path is not None: + cfg_fname = mformer_ckpt2cfg[Path(s1_vfeat_extractor_ckpt_path).name] + else: + cfg_fname = 'divided_224_16x4.yaml' + else: + cfg_fname = 'divided_224_16x4.yaml' + else: + raise ValueError(f'ckpt_path {self.ckpt_path} is not supported.') + else: + was_pt_on_avclip = False + cfg_fname = 'divided_224_16x4.yaml' + # logging.info(f'No ckpt_path provided, using {cfg_fname} config.') + + if cfg_fname in ['motionformer_224_16x4.yaml', 'divided_224_16x4.yaml']: + pos_emb_type = 'separate' + elif cfg_fname == 'joint_224_16x4.yaml': + pos_emb_type = 'joint' + + self.mformer_cfg_path = Path(__file__).absolute().parent / cfg_fname + + check_if_file_exists_else_download(self.mformer_cfg_path, FILE2URL) + mformer_cfg = OmegaConf.load(self.mformer_cfg_path) + logging.info(f'Loading MotionFormer config from {self.mformer_cfg_path.absolute()}') + + # patch the cfg (from the default cfg defined in the repo `Motionformer/slowfast/config/defaults.py`) + mformer_cfg.VIT.ATTN_DROPOUT = 0.0 + mformer_cfg.VIT.POS_EMBED = pos_emb_type + mformer_cfg.VIT.USE_ORIGINAL_TRAJ_ATTN_CODE = True + mformer_cfg.VIT.APPROX_ATTN_TYPE = 'none' # guessing + mformer_cfg.VIT.APPROX_ATTN_DIM = 64 # from ckpt['cfg'] + + # finally init VisionTransformer with the cfg + super().__init__(mformer_cfg) + + # load the ckpt now if ckpt is provided and not from AVCLIPMoCo-pretrained ckpt + if (self.ckpt_path is not None) and (not was_pt_on_avclip): + _ckpt_load_status = self.load_state_dict(ckpt['model_state'], strict=False) + if len(_ckpt_load_status.missing_keys) > 0 or len( + _ckpt_load_status.unexpected_keys) > 0: + logging.warning(f'Loading exact vfeat_extractor ckpt from {self.ckpt_path} failed.' \ + f'Missing keys: {_ckpt_load_status.missing_keys}, ' \ + f'Unexpected keys: {_ckpt_load_status.unexpected_keys}') + else: + logging.info(f'Loading vfeat_extractor ckpt from {self.ckpt_path} succeeded.') + + if self.extract_features: + assert isinstance(self.norm, + nn.LayerNorm), 'early x[:, 1:, :] may not be safe for per-tr weights' + # pre-logits are Sequential(nn.Linear(emb, emd), act) and `act` is tanh but see the logger + self.pre_logits = nn.Identity() + # we don't need the classification head (saving memory) + self.head = nn.Identity() + self.head_drop = nn.Identity() + # avoiding code duplication (used only if agg_*_module is TransformerEncoderLayer) + transf_enc_layer_kwargs = dict( + d_model=self.embed_dim, + nhead=self.num_heads, + activation=nn.GELU(), + batch_first=True, + dim_feedforward=self.mlp_ratio * self.embed_dim, + dropout=self.drop_rate, + layer_norm_eps=1e-6, + norm_first=True, + ) + # define adapters if needed + if self.factorize_space_time: + if agg_space_module == 'TransformerEncoderLayer': + self.spatial_attn_agg = SpatialTransformerEncoderLayer( + **transf_enc_layer_kwargs) + elif agg_space_module == 'AveragePooling': + self.spatial_attn_agg = AveragePooling(avg_pattern='BS D t h w -> BS D t', + then_permute_pattern='BS D t -> BS t D') + if agg_time_module == 'TransformerEncoderLayer': + self.temp_attn_agg = TemporalTransformerEncoderLayer(**transf_enc_layer_kwargs) + elif agg_time_module == 'AveragePooling': + self.temp_attn_agg = AveragePooling(avg_pattern='BS t D -> BS D') + elif 'Identity' in agg_time_module: + self.temp_attn_agg = nn.Identity() + # define a global aggregation layer (aggregarate over segments) + self.add_global_repr = add_global_repr + if add_global_repr: + if agg_segments_module == 'TransformerEncoderLayer': + # we can reuse the same layer as for temporal factorization (B, dim_to_agg, D) -> (B, D) + # we need to add pos emb (PE) because previously we added the same PE for each segment + pos_max_len = max_segments if max_segments is not None else 16 # 16 = 10sec//0.64sec + 1 + self.global_attn_agg = TemporalTransformerEncoderLayer( + add_pos_emb=True, + pos_emb_drop=mformer_cfg.VIT.POS_DROPOUT, + pos_max_len=pos_max_len, + **transf_enc_layer_kwargs) + elif agg_segments_module == 'AveragePooling': + self.global_attn_agg = AveragePooling(avg_pattern='B S D -> B D') + + if was_pt_on_avclip: + # we need to filter out the state_dict of the AVCLIP model (has both A and V extractors) + # and keep only the state_dict of the feat extractor + ckpt_weights = dict() + for k, v in ckpt['state_dict'].items(): + if k.startswith(('module.v_encoder.', 'v_encoder.')): + k = k.replace('module.', '').replace('v_encoder.', '') + ckpt_weights[k] = v + _load_status = self.load_state_dict(ckpt_weights, strict=False) + if len(_load_status.missing_keys) > 0 or len(_load_status.unexpected_keys) > 0: + logging.warning(f'Loading exact vfeat_extractor ckpt from {self.ckpt_path} failed. \n' \ + f'Missing keys ({len(_load_status.missing_keys)}): ' \ + f'{_load_status.missing_keys}, \n' \ + f'Unexpected keys ({len(_load_status.unexpected_keys)}): ' \ + f'{_load_status.unexpected_keys} \n' \ + f'temp_attn_agg are expected to be missing if ckpt was pt contrastively.') + else: + logging.info(f'Loading vfeat_extractor ckpt from {self.ckpt_path} succeeded.') + + # patch_embed is not used in MotionFormer, only patch_embed_3d, because cfg.VIT.PATCH_SIZE_TEMP > 1 + # but it used to calculate the number of patches, so we need to set keep it + self.patch_embed.requires_grad_(False) + + def forward(self, x): + ''' + x is of shape (B, S, C, T, H, W) where S is the number of segments. + ''' + # Batch, Segments, Channels, T=frames, Height, Width + B, S, C, T, H, W = x.shape + # Motionformer expects a tensor of shape (1, B, C, T, H, W). + # The first dimension (1) is a dummy dimension to make the input tensor and won't be used: + # see `video_model_builder.video_input`. + # x = x.unsqueeze(0) # (1, B, S, C, T, H, W) + + orig_shape = (B, S, C, T, H, W) + x = x.view(B * S, C, T, H, W) # flatten batch and segments + x = self.forward_segments(x, orig_shape=orig_shape) + # unpack the segments (using rest dimensions to support different shapes e.g. (BS, D) or (BS, t, D)) + x = x.view(B, S, *x.shape[1:]) + # x is now of shape (B*S, D) or (B*S, t, D) if `self.temp_attn_agg` is `Identity` + + return x # x is (B, S, ...) + + def forward_segments(self, x, orig_shape: tuple) -> torch.Tensor: + '''x is of shape (1, BS, C, T, H, W) where S is the number of segments.''' + x, x_mask = self.forward_features(x) + + assert self.extract_features + + # (BS, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8 + x = x[:, + 1:, :] # without the CLS token for efficiency (should be safe for LayerNorm and FC) + x = self.norm(x) + x = self.pre_logits(x) + if self.factorize_space_time: + x = self.restore_spatio_temp_dims(x, orig_shape) # (B*S, D, t, h, w) <- (B*S, t*h*w, D) + + x = self.spatial_attn_agg(x, x_mask) # (B*S, t, D) + x = self.temp_attn_agg( + x) # (B*S, D) or (BS, t, D) if `self.temp_attn_agg` is `Identity` + + return x + + def restore_spatio_temp_dims(self, feats: torch.Tensor, orig_shape: tuple) -> torch.Tensor: + ''' + feats are of shape (B*S, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8 + Our goal is to make them of shape (B*S, t, h, w, D) where h, w are the spatial dimensions. + From `self.patch_embed_3d`, it follows that we could reshape feats with: + `feats.transpose(1, 2).view(B*S, D, t, h, w)` + ''' + B, S, C, T, H, W = orig_shape + D = self.embed_dim + + # num patches in each dimension + t = T // self.patch_embed_3d.z_block_size + h = self.patch_embed_3d.height + w = self.patch_embed_3d.width + + feats = feats.permute(0, 2, 1) # (B*S, D, T) + feats = feats.view(B * S, D, t, h, w) # (B*S, D, t, h, w) + + return feats + + +class BaseEncoderLayer(nn.TransformerEncoderLayer): + ''' + This is a wrapper around nn.TransformerEncoderLayer that adds a CLS token + to the sequence and outputs the CLS token's representation. + This base class parents both SpatialEncoderLayer and TemporalEncoderLayer for the RGB stream + and the FrequencyEncoderLayer and TemporalEncoderLayer for the audio stream stream. + We also, optionally, add a positional embedding to the input sequence which + allows to reuse it for global aggregation (of segments) for both streams. + ''' + + def __init__(self, + add_pos_emb: bool = False, + pos_emb_drop: float = None, + pos_max_len: int = None, + *args_transformer_enc, + **kwargs_transformer_enc): + super().__init__(*args_transformer_enc, **kwargs_transformer_enc) + self.cls_token = nn.Parameter(torch.zeros(1, 1, self.self_attn.embed_dim)) + trunc_normal_(self.cls_token, std=.02) + + # add positional embedding + self.add_pos_emb = add_pos_emb + if add_pos_emb: + self.pos_max_len = 1 + pos_max_len # +1 (for CLS) + self.pos_emb = nn.Parameter(torch.zeros(1, self.pos_max_len, self.self_attn.embed_dim)) + self.pos_drop = nn.Dropout(pos_emb_drop) + trunc_normal_(self.pos_emb, std=.02) + + self.apply(self._init_weights) + + def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None): + ''' x is of shape (B, N, D); if provided x_mask is of shape (B, N)''' + batch_dim = x.shape[0] + + # add CLS token + cls_tokens = self.cls_token.expand(batch_dim, -1, -1) # expanding to match batch dimension + x = torch.cat((cls_tokens, x), dim=-2) # (batch_dim, 1+seq_len, D) + if x_mask is not None: + cls_mask = torch.ones((batch_dim, 1), dtype=torch.bool, + device=x_mask.device) # 1=keep; 0=mask + x_mask_w_cls = torch.cat((cls_mask, x_mask), dim=-1) # (batch_dim, 1+seq_len) + B, N = x_mask_w_cls.shape + # torch expects (N, N) or (B*num_heads, N, N) mask (sadness ahead); torch masks + x_mask_w_cls = x_mask_w_cls.reshape(B, 1, 1, N)\ + .expand(-1, self.self_attn.num_heads, N, -1)\ + .reshape(B * self.self_attn.num_heads, N, N) + assert x_mask_w_cls.dtype == x_mask_w_cls.bool().dtype, 'x_mask_w_cls.dtype != bool' + x_mask_w_cls = ~x_mask_w_cls # invert mask (1=mask) + else: + x_mask_w_cls = None + + # add positional embedding + if self.add_pos_emb: + seq_len = x.shape[ + 1] # (don't even think about moving it before the CLS token concatenation) + assert seq_len <= self.pos_max_len, f'Seq len ({seq_len}) > pos_max_len ({self.pos_max_len})' + x = x + self.pos_emb[:, :seq_len, :] + x = self.pos_drop(x) + + # apply encoder layer (calls nn.TransformerEncoderLayer.forward); + x = super().forward(src=x, src_mask=x_mask_w_cls) # (batch_dim, 1+seq_len, D) + + # CLS token is expected to hold spatial information for each frame + x = x[:, 0, :] # (batch_dim, D) + + return x + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + return {'cls_token', 'pos_emb'} + + +class SpatialTransformerEncoderLayer(BaseEncoderLayer): + ''' Aggregates spatial dimensions by applying attention individually to each frame. ''' + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None) -> torch.Tensor: + ''' x is of shape (B*S, D, t, h, w) where S is the number of segments. + if specified x_mask (B*S, t, h, w), 0=masked, 1=kept + Returns a tensor of shape (B*S, t, D) pooling spatial information for each frame. ''' + BS, D, t, h, w = x.shape + + # time as a batch dimension and flatten spatial dimensions as sequence + x = einops.rearrange(x, 'BS D t h w -> (BS t) (h w) D') + # similar to mask + if x_mask is not None: + x_mask = einops.rearrange(x_mask, 'BS t h w -> (BS t) (h w)') + + # apply encoder layer (BaseEncoderLayer.forward) - it will add CLS token and output its representation + x = super().forward(x=x, x_mask=x_mask) # (B*S*t, D) + + # reshape back to (B*S, t, D) + x = einops.rearrange(x, '(BS t) D -> BS t D', BS=BS, t=t) + + # (B*S, t, D) + return x + + +class TemporalTransformerEncoderLayer(BaseEncoderLayer): + ''' Aggregates temporal dimension with attention. Also used with pos emb as global aggregation + in both streams. ''' + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, x): + ''' x is of shape (B*S, t, D) where S is the number of segments. + Returns a tensor of shape (B*S, D) pooling temporal information. ''' + BS, t, D = x.shape + + # apply encoder layer (BaseEncoderLayer.forward) - it will add CLS token and output its representation + x = super().forward(x) # (B*S, D) + + return x # (B*S, D) + + +class AveragePooling(nn.Module): + + def __init__(self, avg_pattern: str, then_permute_pattern: str = None) -> None: + ''' patterns are e.g. "bs t d -> bs d" ''' + super().__init__() + # TODO: need to register them as buffers (but fails because these are strings) + self.reduce_fn = 'mean' + self.avg_pattern = avg_pattern + self.then_permute_pattern = then_permute_pattern + + def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None) -> torch.Tensor: + x = einops.reduce(x, self.avg_pattern, self.reduce_fn) + if self.then_permute_pattern is not None: + x = einops.rearrange(x, self.then_permute_pattern) + return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/synchformer.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/synchformer.py new file mode 100644 index 0000000000000000000000000000000000000000..43767156669af3d48f7b75167dcb50f54d560304 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/synchformer.py @@ -0,0 +1,55 @@ +import logging +from typing import Any, Mapping + +import torch +from torch import nn + +from ...ext.synchformer.motionformer import MotionFormer + + +class Synchformer(nn.Module): + + def __init__(self): + super().__init__() + + self.vfeat_extractor = MotionFormer(extract_features=True, + factorize_space_time=True, + agg_space_module='TransformerEncoderLayer', + agg_time_module='torch.nn.Identity', + add_global_repr=False) + + # self.vfeat_extractor = instantiate_from_config(vfeat_extractor) + # self.afeat_extractor = instantiate_from_config(afeat_extractor) + # # bridging the s3d latent dim (1024) into what is specified in the config + # # to match e.g. the transformer dim + # self.vproj = instantiate_from_config(vproj) + # self.aproj = instantiate_from_config(aproj) + # self.transformer = instantiate_from_config(transformer) + + def forward(self, vis): + B, S, Tv, C, H, W = vis.shape + vis = vis.permute(0, 1, 3, 2, 4, 5) # (B, S, C, Tv, H, W) + # feat extractors return a tuple of segment-level and global features (ignored for sync) + # (B, S, tv, D), e.g. (B, 7, 8, 768) + vis = self.vfeat_extractor(vis) + return vis + + def load_state_dict(self, sd: Mapping[str, Any], strict: bool = True): + # discard all entries except vfeat_extractor + sd = {k: v for k, v in sd.items() if k.startswith('vfeat_extractor')} + + return super().load_state_dict(sd, strict) + + +if __name__ == "__main__": + model = Synchformer().cuda().eval() + sd = torch.load('./ext_weights/synchformer_state_dict.pth', weights_only=True) + model.load_state_dict(sd) + + vid = torch.randn(2, 7, 16, 3, 224, 224).cuda() + features = model.extract_vfeats(vid, for_loop=False).detach().cpu() + print(features.shape) + + # extract and save the state dict only + # sd = torch.load('./ext_weights/sync_model_audioset.pt')['model'] + # torch.save(sd, './ext_weights/synchformer_state_dict.pth') diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a797eb9c66f04b7c29934bfc384c935cdf441a62 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/utils.py @@ -0,0 +1,92 @@ +from hashlib import md5 +from pathlib import Path + +import requests +from tqdm import tqdm + +PARENT_LINK = 'https://a3s.fi/swift/v1/AUTH_a235c0f452d648828f745589cde1219a' +FNAME2LINK = { + # S3: Synchability: AudioSet (run 2) + '24-01-22T20-34-52.pt': + f'{PARENT_LINK}/sync/sync_models/24-01-22T20-34-52/24-01-22T20-34-52.pt', + 'cfg-24-01-22T20-34-52.yaml': + f'{PARENT_LINK}/sync/sync_models/24-01-22T20-34-52/cfg-24-01-22T20-34-52.yaml', + # S2: Synchformer: AudioSet (run 2) + '24-01-04T16-39-21.pt': + f'{PARENT_LINK}/sync/sync_models/24-01-04T16-39-21/24-01-04T16-39-21.pt', + 'cfg-24-01-04T16-39-21.yaml': + f'{PARENT_LINK}/sync/sync_models/24-01-04T16-39-21/cfg-24-01-04T16-39-21.yaml', + # S2: Synchformer: AudioSet (run 1) + '23-08-28T11-23-23.pt': + f'{PARENT_LINK}/sync/sync_models/23-08-28T11-23-23/23-08-28T11-23-23.pt', + 'cfg-23-08-28T11-23-23.yaml': + f'{PARENT_LINK}/sync/sync_models/23-08-28T11-23-23/cfg-23-08-28T11-23-23.yaml', + # S2: Synchformer: LRS3 (run 2) + '23-12-23T18-33-57.pt': + f'{PARENT_LINK}/sync/sync_models/23-12-23T18-33-57/23-12-23T18-33-57.pt', + 'cfg-23-12-23T18-33-57.yaml': + f'{PARENT_LINK}/sync/sync_models/23-12-23T18-33-57/cfg-23-12-23T18-33-57.yaml', + # S2: Synchformer: VGS (run 2) + '24-01-02T10-00-53.pt': + f'{PARENT_LINK}/sync/sync_models/24-01-02T10-00-53/24-01-02T10-00-53.pt', + 'cfg-24-01-02T10-00-53.yaml': + f'{PARENT_LINK}/sync/sync_models/24-01-02T10-00-53/cfg-24-01-02T10-00-53.yaml', + # SparseSync: ft VGGSound-Full + '22-09-21T21-00-52.pt': + f'{PARENT_LINK}/sync/sync_models/22-09-21T21-00-52/22-09-21T21-00-52.pt', + 'cfg-22-09-21T21-00-52.yaml': + f'{PARENT_LINK}/sync/sync_models/22-09-21T21-00-52/cfg-22-09-21T21-00-52.yaml', + # SparseSync: ft VGGSound-Sparse + '22-07-28T15-49-45.pt': + f'{PARENT_LINK}/sync/sync_models/22-07-28T15-49-45/22-07-28T15-49-45.pt', + 'cfg-22-07-28T15-49-45.yaml': + f'{PARENT_LINK}/sync/sync_models/22-07-28T15-49-45/cfg-22-07-28T15-49-45.yaml', + # SparseSync: only pt on LRS3 + '22-07-13T22-25-49.pt': + f'{PARENT_LINK}/sync/sync_models/22-07-13T22-25-49/22-07-13T22-25-49.pt', + 'cfg-22-07-13T22-25-49.yaml': + f'{PARENT_LINK}/sync/sync_models/22-07-13T22-25-49/cfg-22-07-13T22-25-49.yaml', + # SparseSync: feature extractors + 'ResNetAudio-22-08-04T09-51-04.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-08-04T09-51-04.pt', # 2s + 'ResNetAudio-22-08-03T23-14-49.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-08-03T23-14-49.pt', # 3s + 'ResNetAudio-22-08-03T23-14-28.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-08-03T23-14-28.pt', # 4s + 'ResNetAudio-22-06-24T08-10-33.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-06-24T08-10-33.pt', # 5s + 'ResNetAudio-22-06-24T17-31-07.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-06-24T17-31-07.pt', # 6s + 'ResNetAudio-22-06-24T23-57-11.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-06-24T23-57-11.pt', # 7s + 'ResNetAudio-22-06-25T04-35-42.pt': + f'{PARENT_LINK}/sync/ResNetAudio-22-06-25T04-35-42.pt', # 8s +} + + +def check_if_file_exists_else_download(path, fname2link=FNAME2LINK, chunk_size=1024): + '''Checks if file exists, if not downloads it from the link to the path''' + path = Path(path) + if not path.exists(): + path.parent.mkdir(exist_ok=True, parents=True) + link = fname2link.get(path.name, None) + if link is None: + raise ValueError(f'Cant find the checkpoint file: {path}.', + f'Please download it manually and ensure the path exists.') + with requests.get(fname2link[path.name], stream=True) as r: + total_size = int(r.headers.get('content-length', 0)) + with tqdm(total=total_size, unit='B', unit_scale=True) as pbar: + with open(path, 'wb') as f: + for data in r.iter_content(chunk_size=chunk_size): + if data: + f.write(data) + pbar.update(chunk_size) + + +def get_md5sum(path): + hash_md5 = md5() + with open(path, 'rb') as f: + for chunk in iter(lambda: f.read(4096 * 8), b''): + hash_md5.update(chunk) + md5sum = hash_md5.hexdigest() + return md5sum diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/video_model_builder.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/video_model_builder.py new file mode 100644 index 0000000000000000000000000000000000000000..9b3582fa956046c0f020fdde9c2c6bd3a0aa1f8d --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/video_model_builder.py @@ -0,0 +1,277 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. +# Copyright 2020 Ross Wightman +# Modified Model definition + +from collections import OrderedDict +from functools import partial + +import torch +import torch.nn as nn +from timm.layers import trunc_normal_ + +from ...ext.synchformer import vit_helper + + +class VisionTransformer(nn.Module): + """ Vision Transformer with support for patch or hybrid CNN input stage """ + + def __init__(self, cfg): + super().__init__() + self.img_size = cfg.DATA.TRAIN_CROP_SIZE + self.patch_size = cfg.VIT.PATCH_SIZE + self.in_chans = cfg.VIT.CHANNELS + if cfg.TRAIN.DATASET == "Epickitchens": + self.num_classes = [97, 300] + else: + self.num_classes = cfg.MODEL.NUM_CLASSES + self.embed_dim = cfg.VIT.EMBED_DIM + self.depth = cfg.VIT.DEPTH + self.num_heads = cfg.VIT.NUM_HEADS + self.mlp_ratio = cfg.VIT.MLP_RATIO + self.qkv_bias = cfg.VIT.QKV_BIAS + self.drop_rate = cfg.VIT.DROP + self.drop_path_rate = cfg.VIT.DROP_PATH + self.head_dropout = cfg.VIT.HEAD_DROPOUT + self.video_input = cfg.VIT.VIDEO_INPUT + self.temporal_resolution = cfg.VIT.TEMPORAL_RESOLUTION + self.use_mlp = cfg.VIT.USE_MLP + self.num_features = self.embed_dim + norm_layer = partial(nn.LayerNorm, eps=1e-6) + self.attn_drop_rate = cfg.VIT.ATTN_DROPOUT + self.head_act = cfg.VIT.HEAD_ACT + self.cfg = cfg + + # Patch Embedding + self.patch_embed = vit_helper.PatchEmbed(img_size=224, + patch_size=self.patch_size, + in_chans=self.in_chans, + embed_dim=self.embed_dim) + + # 3D Patch Embedding + self.patch_embed_3d = vit_helper.PatchEmbed3D(img_size=self.img_size, + temporal_resolution=self.temporal_resolution, + patch_size=self.patch_size, + in_chans=self.in_chans, + embed_dim=self.embed_dim, + z_block_size=self.cfg.VIT.PATCH_SIZE_TEMP) + self.patch_embed_3d.proj.weight.data = torch.zeros_like( + self.patch_embed_3d.proj.weight.data) + + # Number of patches + if self.video_input: + num_patches = self.patch_embed.num_patches * self.temporal_resolution + else: + num_patches = self.patch_embed.num_patches + self.num_patches = num_patches + + # CLS token + self.cls_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim)) + trunc_normal_(self.cls_token, std=.02) + + # Positional embedding + self.pos_embed = nn.Parameter( + torch.zeros(1, self.patch_embed.num_patches + 1, self.embed_dim)) + self.pos_drop = nn.Dropout(p=cfg.VIT.POS_DROPOUT) + trunc_normal_(self.pos_embed, std=.02) + + if self.cfg.VIT.POS_EMBED == "joint": + self.st_embed = nn.Parameter(torch.zeros(1, num_patches + 1, self.embed_dim)) + trunc_normal_(self.st_embed, std=.02) + elif self.cfg.VIT.POS_EMBED == "separate": + self.temp_embed = nn.Parameter(torch.zeros(1, self.temporal_resolution, self.embed_dim)) + + # Layer Blocks + dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, self.depth)] + if self.cfg.VIT.ATTN_LAYER == "divided": + self.blocks = nn.ModuleList([ + vit_helper.DividedSpaceTimeBlock( + attn_type=cfg.VIT.ATTN_LAYER, + dim=self.embed_dim, + num_heads=self.num_heads, + mlp_ratio=self.mlp_ratio, + qkv_bias=self.qkv_bias, + drop=self.drop_rate, + attn_drop=self.attn_drop_rate, + drop_path=dpr[i], + norm_layer=norm_layer, + ) for i in range(self.depth) + ]) + else: + self.blocks = nn.ModuleList([ + vit_helper.Block(attn_type=cfg.VIT.ATTN_LAYER, + dim=self.embed_dim, + num_heads=self.num_heads, + mlp_ratio=self.mlp_ratio, + qkv_bias=self.qkv_bias, + drop=self.drop_rate, + attn_drop=self.attn_drop_rate, + drop_path=dpr[i], + norm_layer=norm_layer, + use_original_code=self.cfg.VIT.USE_ORIGINAL_TRAJ_ATTN_CODE) + for i in range(self.depth) + ]) + self.norm = norm_layer(self.embed_dim) + + # MLP head + if self.use_mlp: + hidden_dim = self.embed_dim + if self.head_act == 'tanh': + # logging.info("Using TanH activation in MLP") + act = nn.Tanh() + elif self.head_act == 'gelu': + # logging.info("Using GELU activation in MLP") + act = nn.GELU() + else: + # logging.info("Using ReLU activation in MLP") + act = nn.ReLU() + self.pre_logits = nn.Sequential( + OrderedDict([ + ('fc', nn.Linear(self.embed_dim, hidden_dim)), + ('act', act), + ])) + else: + self.pre_logits = nn.Identity() + + # Classifier Head + self.head_drop = nn.Dropout(p=self.head_dropout) + if isinstance(self.num_classes, (list, )) and len(self.num_classes) > 1: + for a, i in enumerate(range(len(self.num_classes))): + setattr(self, "head%d" % a, nn.Linear(self.embed_dim, self.num_classes[i])) + else: + self.head = nn.Linear(self.embed_dim, + self.num_classes) if self.num_classes > 0 else nn.Identity() + + # Initialize weights + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + if self.cfg.VIT.POS_EMBED == "joint": + return {'pos_embed', 'cls_token', 'st_embed'} + else: + return {'pos_embed', 'cls_token', 'temp_embed'} + + def get_classifier(self): + return self.head + + def reset_classifier(self, num_classes, global_pool=''): + self.num_classes = num_classes + self.head = (nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()) + + def forward_features(self, x): + # if self.video_input: + # x = x[0] + B = x.shape[0] + + # Tokenize input + # if self.cfg.VIT.PATCH_SIZE_TEMP > 1: + # for simplicity of mapping between content dimensions (input x) and token dims (after patching) + # we use the same trick as for AST (see modeling_ast.ASTModel.forward for the details): + + # apply patching on input + x = self.patch_embed_3d(x) + tok_mask = None + + # else: + # tok_mask = None + # # 2D tokenization + # if self.video_input: + # x = x.permute(0, 2, 1, 3, 4) + # (B, T, C, H, W) = x.shape + # x = x.reshape(B * T, C, H, W) + + # x = self.patch_embed(x) + + # if self.video_input: + # (B2, T2, D2) = x.shape + # x = x.reshape(B, T * T2, D2) + + # Append CLS token + cls_tokens = self.cls_token.expand(B, -1, -1) + x = torch.cat((cls_tokens, x), dim=1) + # if tok_mask is not None: + # # prepend 1(=keep) to the mask to account for the CLS token as well + # tok_mask = torch.cat((torch.ones_like(tok_mask[:, [0]]), tok_mask), dim=1) + + # Interpolate positinoal embeddings + # if self.cfg.DATA.TRAIN_CROP_SIZE != 224: + # pos_embed = self.pos_embed + # N = pos_embed.shape[1] - 1 + # npatch = int((x.size(1) - 1) / self.temporal_resolution) + # class_emb = pos_embed[:, 0] + # pos_embed = pos_embed[:, 1:] + # dim = x.shape[-1] + # pos_embed = torch.nn.functional.interpolate( + # pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2), + # scale_factor=math.sqrt(npatch / N), + # mode='bicubic', + # ) + # pos_embed = pos_embed.permute(0, 2, 3, 1).view(1, -1, dim) + # new_pos_embed = torch.cat((class_emb.unsqueeze(0), pos_embed), dim=1) + # else: + new_pos_embed = self.pos_embed + npatch = self.patch_embed.num_patches + + # Add positional embeddings to input + if self.video_input: + if self.cfg.VIT.POS_EMBED == "separate": + cls_embed = self.pos_embed[:, 0, :].unsqueeze(1) + tile_pos_embed = new_pos_embed[:, 1:, :].repeat(1, self.temporal_resolution, 1) + tile_temporal_embed = self.temp_embed.repeat_interleave(npatch, 1) + total_pos_embed = tile_pos_embed + tile_temporal_embed + total_pos_embed = torch.cat([cls_embed, total_pos_embed], dim=1) + x = x + total_pos_embed + elif self.cfg.VIT.POS_EMBED == "joint": + x = x + self.st_embed + else: + # image input + x = x + new_pos_embed + + # Apply positional dropout + x = self.pos_drop(x) + + # Encoding using transformer layers + for i, blk in enumerate(self.blocks): + x = blk(x, + seq_len=npatch, + num_frames=self.temporal_resolution, + approx=self.cfg.VIT.APPROX_ATTN_TYPE, + num_landmarks=self.cfg.VIT.APPROX_ATTN_DIM, + tok_mask=tok_mask) + + ### v-iashin: I moved it to the forward pass + # x = self.norm(x)[:, 0] + # x = self.pre_logits(x) + ### + return x, tok_mask + + # def forward(self, x): + # x = self.forward_features(x) + # ### v-iashin: here. This should leave the same forward output as before + # x = self.norm(x)[:, 0] + # x = self.pre_logits(x) + # ### + # x = self.head_drop(x) + # if isinstance(self.num_classes, (list, )) and len(self.num_classes) > 1: + # output = [] + # for head in range(len(self.num_classes)): + # x_out = getattr(self, "head%d" % head)(x) + # if not self.training: + # x_out = torch.nn.functional.softmax(x_out, dim=-1) + # output.append(x_out) + # return output + # else: + # x = self.head(x) + # if not self.training: + # x = torch.nn.functional.softmax(x, dim=-1) + # return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/vit_helper.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/vit_helper.py new file mode 100644 index 0000000000000000000000000000000000000000..6af730a135bf49240ec439c81c9ad0aa5c9a505e --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/ext/synchformer/vit_helper.py @@ -0,0 +1,399 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. +# Copyright 2020 Ross Wightman +# Modified Model definition +"""Video models.""" + +import math + +import torch +import torch.nn as nn +from einops import rearrange, repeat +from timm.layers import to_2tuple +from torch import einsum +from torch.nn import functional as F + +default_cfgs = { + 'vit_1k': + 'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth', + 'vit_1k_large': + 'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_224-4ee7a4dc.pth', +} + + +def qkv_attn(q, k, v, tok_mask: torch.Tensor = None): + sim = einsum('b i d, b j d -> b i j', q, k) + # apply masking if provided, tok_mask is (B*S*H, N): 1s - keep; sim is (B*S*H, H, N, N) + if tok_mask is not None: + BSH, N = tok_mask.shape + sim = sim.masked_fill(tok_mask.view(BSH, 1, N) == 0, + float('-inf')) # 1 - broadcasts across N + attn = sim.softmax(dim=-1) + out = einsum('b i j, b j d -> b i d', attn, v) + return out + + +class DividedAttention(nn.Module): + + def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = head_dim**-0.5 + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.proj = nn.Linear(dim, dim) + + # init to zeros + self.qkv.weight.data.fill_(0) + self.qkv.bias.data.fill_(0) + self.proj.weight.data.fill_(1) + self.proj.bias.data.fill_(0) + + self.attn_drop = nn.Dropout(attn_drop) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x, einops_from, einops_to, tok_mask: torch.Tensor = None, **einops_dims): + # num of heads variable + h = self.num_heads + + # project x to q, k, v vaalues + q, k, v = self.qkv(x).chunk(3, dim=-1) + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + if tok_mask is not None: + # replicate token mask across heads (b, n) -> (b, h, n) -> (b*h, n) -- same as qkv but w/o d + assert len(tok_mask.shape) == 2 + tok_mask = tok_mask.unsqueeze(1).expand(-1, h, -1).reshape(-1, tok_mask.shape[1]) + + # Scale q + q *= self.scale + + # Take out cls_q, cls_k, cls_v + (cls_q, q_), (cls_k, k_), (cls_v, v_) = map(lambda t: (t[:, 0:1], t[:, 1:]), (q, k, v)) + # the same for masking + if tok_mask is not None: + cls_mask, mask_ = tok_mask[:, 0:1], tok_mask[:, 1:] + else: + cls_mask, mask_ = None, None + + # let CLS token attend to key / values of all patches across time and space + cls_out = qkv_attn(cls_q, k, v, tok_mask=tok_mask) + + # rearrange across time or space + q_, k_, v_ = map(lambda t: rearrange(t, f'{einops_from} -> {einops_to}', **einops_dims), + (q_, k_, v_)) + + # expand CLS token keys and values across time or space and concat + r = q_.shape[0] // cls_k.shape[0] + cls_k, cls_v = map(lambda t: repeat(t, 'b () d -> (b r) () d', r=r), (cls_k, cls_v)) + + k_ = torch.cat((cls_k, k_), dim=1) + v_ = torch.cat((cls_v, v_), dim=1) + + # the same for masking (if provided) + if tok_mask is not None: + # since mask does not have the latent dim (d), we need to remove it from einops dims + mask_ = rearrange(mask_, f'{einops_from} -> {einops_to}'.replace(' d', ''), + **einops_dims) + cls_mask = repeat(cls_mask, 'b () -> (b r) ()', + r=r) # expand cls_mask across time or space + mask_ = torch.cat((cls_mask, mask_), dim=1) + + # attention + out = qkv_attn(q_, k_, v_, tok_mask=mask_) + + # merge back time or space + out = rearrange(out, f'{einops_to} -> {einops_from}', **einops_dims) + + # concat back the cls token + out = torch.cat((cls_out, out), dim=1) + + # merge back the heads + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + + ## to out + x = self.proj(out) + x = self.proj_drop(x) + return x + + +class DividedSpaceTimeBlock(nn.Module): + + def __init__(self, + dim=768, + num_heads=12, + attn_type='divided', + mlp_ratio=4., + qkv_bias=False, + drop=0., + attn_drop=0., + drop_path=0., + act_layer=nn.GELU, + norm_layer=nn.LayerNorm): + super().__init__() + + self.einops_from_space = 'b (f n) d' + self.einops_to_space = '(b f) n d' + self.einops_from_time = 'b (f n) d' + self.einops_to_time = '(b n) f d' + + self.norm1 = norm_layer(dim) + + self.attn = DividedAttention(dim, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop) + + self.timeattn = DividedAttention(dim, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop) + + # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.drop_path = nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop) + self.norm3 = norm_layer(dim) + + def forward(self, + x, + seq_len=196, + num_frames=8, + approx='none', + num_landmarks=128, + tok_mask: torch.Tensor = None): + time_output = self.timeattn(self.norm3(x), + self.einops_from_time, + self.einops_to_time, + n=seq_len, + tok_mask=tok_mask) + time_residual = x + time_output + + space_output = self.attn(self.norm1(time_residual), + self.einops_from_space, + self.einops_to_space, + f=num_frames, + tok_mask=tok_mask) + space_residual = time_residual + self.drop_path(space_output) + + x = space_residual + x = x + self.drop_path(self.mlp(self.norm2(x))) + return x + + +class Mlp(nn.Module): + + def __init__(self, + in_features, + hidden_features=None, + out_features=None, + act_layer=nn.GELU, + drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + """ + + def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): + super().__init__() + img_size = img_size if type(img_size) is tuple else to_2tuple(img_size) + patch_size = img_size if type(patch_size) is tuple else to_2tuple(patch_size) + num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) + self.img_size = img_size + self.patch_size = patch_size + self.num_patches = num_patches + + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + + def forward(self, x): + B, C, H, W = x.shape + x = self.proj(x).flatten(2).transpose(1, 2) + return x + + +class PatchEmbed3D(nn.Module): + """ Image to Patch Embedding """ + + def __init__(self, + img_size=224, + temporal_resolution=4, + in_chans=3, + patch_size=16, + z_block_size=2, + embed_dim=768, + flatten=True): + super().__init__() + self.height = (img_size // patch_size) + self.width = (img_size // patch_size) + ### v-iashin: these two are incorrect + # self.frames = (temporal_resolution // z_block_size) + # self.num_patches = self.height * self.width * self.frames + self.z_block_size = z_block_size + ### + self.proj = nn.Conv3d(in_chans, + embed_dim, + kernel_size=(z_block_size, patch_size, patch_size), + stride=(z_block_size, patch_size, patch_size)) + self.flatten = flatten + + def forward(self, x): + B, C, T, H, W = x.shape + x = self.proj(x) + if self.flatten: + x = x.flatten(2).transpose(1, 2) + return x + + +class HeadMLP(nn.Module): + + def __init__(self, n_input, n_classes, n_hidden=512, p=0.1): + super(HeadMLP, self).__init__() + self.n_input = n_input + self.n_classes = n_classes + self.n_hidden = n_hidden + if n_hidden is None: + # use linear classifier + self.block_forward = nn.Sequential(nn.Dropout(p=p), + nn.Linear(n_input, n_classes, bias=True)) + else: + # use simple MLP classifier + self.block_forward = nn.Sequential(nn.Dropout(p=p), + nn.Linear(n_input, n_hidden, bias=True), + nn.BatchNorm1d(n_hidden), nn.ReLU(inplace=True), + nn.Dropout(p=p), + nn.Linear(n_hidden, n_classes, bias=True)) + print(f"Dropout-NLP: {p}") + + def forward(self, x): + return self.block_forward(x) + + +def _conv_filter(state_dict, patch_size=16): + """ convert patch embedding weight from manual patchify + linear proj to conv""" + out_dict = {} + for k, v in state_dict.items(): + if 'patch_embed.proj.weight' in k: + v = v.reshape((v.shape[0], 3, patch_size, patch_size)) + out_dict[k] = v + return out_dict + + +def adapt_input_conv(in_chans, conv_weight, agg='sum'): + conv_type = conv_weight.dtype + conv_weight = conv_weight.float() + O, I, J, K = conv_weight.shape + if in_chans == 1: + if I > 3: + assert conv_weight.shape[1] % 3 == 0 + # For models with space2depth stems + conv_weight = conv_weight.reshape(O, I // 3, 3, J, K) + conv_weight = conv_weight.sum(dim=2, keepdim=False) + else: + if agg == 'sum': + print("Summing conv1 weights") + conv_weight = conv_weight.sum(dim=1, keepdim=True) + else: + print("Averaging conv1 weights") + conv_weight = conv_weight.mean(dim=1, keepdim=True) + elif in_chans != 3: + if I != 3: + raise NotImplementedError('Weight format not supported by conversion.') + else: + if agg == 'sum': + print("Summing conv1 weights") + repeat = int(math.ceil(in_chans / 3)) + conv_weight = conv_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :] + conv_weight *= (3 / float(in_chans)) + else: + print("Averaging conv1 weights") + conv_weight = conv_weight.mean(dim=1, keepdim=True) + conv_weight = conv_weight.repeat(1, in_chans, 1, 1) + conv_weight = conv_weight.to(conv_type) + return conv_weight + + +def load_pretrained(model, + cfg=None, + num_classes=1000, + in_chans=3, + filter_fn=None, + strict=True, + progress=False): + # Load state dict + assert (f"{cfg.VIT.PRETRAINED_WEIGHTS} not in [vit_1k, vit_1k_large]") + state_dict = torch.hub.load_state_dict_from_url(url=default_cfgs[cfg.VIT.PRETRAINED_WEIGHTS]) + + if filter_fn is not None: + state_dict = filter_fn(state_dict) + + input_convs = 'patch_embed.proj' + if input_convs is not None and in_chans != 3: + if isinstance(input_convs, str): + input_convs = (input_convs, ) + for input_conv_name in input_convs: + weight_name = input_conv_name + '.weight' + try: + state_dict[weight_name] = adapt_input_conv(in_chans, + state_dict[weight_name], + agg='avg') + print( + f'Converted input conv {input_conv_name} pretrained weights from 3 to {in_chans} channel(s)' + ) + except NotImplementedError as e: + del state_dict[weight_name] + strict = False + print( + f'Unable to convert pretrained {input_conv_name} weights, using random init for this layer.' + ) + + classifier_name = 'head' + label_offset = cfg.get('label_offset', 0) + pretrain_classes = 1000 + if num_classes != pretrain_classes: + # completely discard fully connected if model num_classes doesn't match pretrained weights + del state_dict[classifier_name + '.weight'] + del state_dict[classifier_name + '.bias'] + strict = False + elif label_offset > 0: + # special case for pretrained weights with an extra background class in pretrained weights + classifier_weight = state_dict[classifier_name + '.weight'] + state_dict[classifier_name + '.weight'] = classifier_weight[label_offset:] + classifier_bias = state_dict[classifier_name + '.bias'] + state_dict[classifier_name + '.bias'] = classifier_bias[label_offset:] + + loaded_state = state_dict + self_state = model.state_dict() + all_names = set(self_state.keys()) + saved_names = set([]) + for name, param in loaded_state.items(): + param = param + if 'module.' in name: + name = name.replace('module.', '') + if name in self_state.keys() and param.shape == self_state[name].shape: + saved_names.add(name) + self_state[name].copy_(param) + else: + print(f"didnt load: {name} of shape: {param.shape}") + print("Missing Keys:") + print(all_names - saved_names) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/embeddings.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/embeddings.py new file mode 100644 index 0000000000000000000000000000000000000000..297feb4d2c79d306771f5436dbd4ada1a976b3bc --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/embeddings.py @@ -0,0 +1,49 @@ +import torch +import torch.nn as nn + +# https://github.com/facebookresearch/DiT + + +class TimestepEmbedder(nn.Module): + """ + Embeds scalar timesteps into vector representations. + """ + + def __init__(self, dim, frequency_embedding_size, max_period): + super().__init__() + self.mlp = nn.Sequential( + nn.Linear(frequency_embedding_size, dim), + nn.SiLU(), + nn.Linear(dim, dim), + ) + self.dim = dim + self.max_period = max_period + assert dim % 2 == 0, 'dim must be even.' + + with torch.autocast('cuda', enabled=False): + self.freqs = nn.Buffer( + 1.0 / (10000**(torch.arange(0, frequency_embedding_size, 2, dtype=torch.float32) / + frequency_embedding_size)), + persistent=False) + freq_scale = 10000 / max_period + self.freqs = freq_scale * self.freqs + + def timestep_embedding(self, t): + """ + Create sinusoidal timestep embeddings. + :param t: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an (N, D) Tensor of positional embeddings. + """ + # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py + + args = t[:, None].float() * self.freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + return embedding + + def forward(self, t): + t_freq = self.timestep_embedding(t).to(t.dtype) + t_emb = self.mlp(t_freq) + return t_emb diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/flow_matching.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/flow_matching.py new file mode 100644 index 0000000000000000000000000000000000000000..d3f491a3fa47bcba0a6eeb1cebdc1eec4258a6cd --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/flow_matching.py @@ -0,0 +1,93 @@ +import logging +from typing import Callable, Iterable, Optional + +import torch +from torchdiffeq import odeint + +# from torchcfm.conditional_flow_matching import ExactOptimalTransportConditionalFlowMatcher + +log = logging.getLogger() +from comfy.utils import ProgressBar +from tqdm import tqdm + +# Partially from https://github.com/gle-bellier/flow-matching +class FlowMatching: + + def __init__(self, min_sigma: float = 0.0, inference_mode='euler', num_steps: int = 25): + # inference_mode: 'euler' or 'adaptive' + # num_steps: number of steps in the euler inference mode + super().__init__() + self.min_sigma = min_sigma + self.inference_mode = inference_mode + self.num_steps = num_steps + + # self.fm = ExactOptimalTransportConditionalFlowMatcher(sigma=min_sigma) + + assert self.inference_mode in ['euler', 'adaptive'] + if self.inference_mode == 'adaptive' and num_steps > 0: + log.info('The number of steps is ignored in adaptive inference mode ') + + def get_conditional_flow(self, x0: torch.Tensor, x1: torch.Tensor, + t: torch.Tensor) -> torch.Tensor: + # which is psi_t(x), eq 22 in flow matching for generative models + t = t[:, None, None].expand_as(x0) + return (1 - (1 - self.min_sigma) * t) * x0 + t * x1 + + def loss(self, predicted_v: torch.Tensor, x0: torch.Tensor, x1: torch.Tensor) -> torch.Tensor: + # return the mean error without reducing the batch dimension + reduce_dim = list(range(1, len(predicted_v.shape))) + target_v = x1 - (1 - self.min_sigma) * x0 + return (predicted_v - target_v).pow(2).mean(dim=reduce_dim) + + def get_x0_xt_c( + self, + x1: torch.Tensor, + t: torch.Tensor, + Cs: list[torch.Tensor], + generator: Optional[torch.Generator] = None + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + # x0 = torch.randn_like(x1, generator=generator) + x0 = torch.empty_like(x1).normal_(generator=generator) + + # find mini-batch optimal transport + # x0, x1, _, Cs = self.fm.ot_sampler.sample_plan_with_labels(x0, x1, None, Cs, replace=True) + + xt = self.get_conditional_flow(x0, x1, t) + return x0, x1, xt, Cs + + def to_prior(self, fn: Callable, x1: torch.Tensor) -> torch.Tensor: + return self.run_t0_to_t1(fn, x1, 1, 0) + + def to_data(self, fn: Callable, x0: torch.Tensor) -> torch.Tensor: + return self.run_t0_to_t1(fn, x0, 0, 1) + + def run_t0_to_t1(self, fn: Callable, x0: torch.Tensor, t0: float, t1: float) -> torch.Tensor: + # fn: a function that takes (t, x) and returns the direction x0->x1 + + if self.inference_mode == 'adaptive': + return odeint(fn, x0, torch.tensor([t0, t1], device=x0.device, dtype=x0.dtype)) + elif self.inference_mode == 'euler': + x = x0 + steps = torch.linspace(t0, t1 - self.min_sigma, self.num_steps + 1) + comfy_pbar = ProgressBar(len(steps)-1) + tqdm_pbar = tqdm(total=len(steps) -1, desc='Flow Matching') + for ti, t in enumerate(steps[:-1]): + flow = fn(t, x) + next_t = steps[ti + 1] + dt = next_t - t + x = x + dt * flow + comfy_pbar.update(1) + tqdm_pbar.update(1) + + # return odeint(fn, + # x0, + # torch.tensor([t0, t1], device=x0.device, dtype=x0.dtype), + # method='rk4', + # options=dict(step_size=(t1 - t0) / self.num_steps))[-1] + # return odeint(fn, + # x0, + # torch.tensor([t0, t1], device=x0.device, dtype=x0.dtype), + # method='euler', + # options=dict(step_size=(t1 - t0) / self.num_steps))[-1] + + return x diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/low_level.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/low_level.py new file mode 100644 index 0000000000000000000000000000000000000000..c8326a8bec99f1be08b92e76fda4b59e777b39d2 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/low_level.py @@ -0,0 +1,95 @@ +import torch +from torch import nn +from torch.nn import functional as F + + +class ChannelLastConv1d(nn.Conv1d): + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = x.permute(0, 2, 1) + x = super().forward(x) + x = x.permute(0, 2, 1) + return x + + +# https://github.com/Stability-AI/sd3-ref +class MLP(nn.Module): + + def __init__( + self, + dim: int, + hidden_dim: int, + multiple_of: int = 256, + ): + """ + Initialize the FeedForward module. + + Args: + dim (int): Input dimension. + hidden_dim (int): Hidden dimension of the feedforward layer. + multiple_of (int): Value to ensure hidden dimension is a multiple of this value. + + Attributes: + w1 (ColumnParallelLinear): Linear transformation for the first layer. + w2 (RowParallelLinear): Linear transformation for the second layer. + w3 (ColumnParallelLinear): Linear transformation for the third layer. + + """ + super().__init__() + hidden_dim = int(2 * hidden_dim / 3) + hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) + + self.w1 = nn.Linear(dim, hidden_dim, bias=False) + self.w2 = nn.Linear(hidden_dim, dim, bias=False) + self.w3 = nn.Linear(dim, hidden_dim, bias=False) + + def forward(self, x): + return self.w2(F.silu(self.w1(x)) * self.w3(x)) + + +class ConvMLP(nn.Module): + + def __init__( + self, + dim: int, + hidden_dim: int, + multiple_of: int = 256, + kernel_size: int = 3, + padding: int = 1, + ): + """ + Initialize the FeedForward module. + + Args: + dim (int): Input dimension. + hidden_dim (int): Hidden dimension of the feedforward layer. + multiple_of (int): Value to ensure hidden dimension is a multiple of this value. + + Attributes: + w1 (ColumnParallelLinear): Linear transformation for the first layer. + w2 (RowParallelLinear): Linear transformation for the second layer. + w3 (ColumnParallelLinear): Linear transformation for the third layer. + + """ + super().__init__() + hidden_dim = int(2 * hidden_dim / 3) + hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) + + self.w1 = ChannelLastConv1d(dim, + hidden_dim, + bias=False, + kernel_size=kernel_size, + padding=padding) + self.w2 = ChannelLastConv1d(hidden_dim, + dim, + bias=False, + kernel_size=kernel_size, + padding=padding) + self.w3 = ChannelLastConv1d(dim, + hidden_dim, + bias=False, + kernel_size=kernel_size, + padding=padding) + + def forward(self, x): + return self.w2(F.silu(self.w1(x)) * self.w3(x)) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/networks.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/networks.py new file mode 100644 index 0000000000000000000000000000000000000000..76d2fe1439e3688afeb28e5320b16d51fd4f5655 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/networks.py @@ -0,0 +1,469 @@ +import logging +from dataclasses import dataclass +from typing import Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ext.rotary_embeddings import compute_rope_rotations +from .embeddings import TimestepEmbedder +from .low_level import MLP, ChannelLastConv1d, ConvMLP +from .transformer_layers import (FinalBlock, JointBlock, MMDitSingleBlock) + +log = logging.getLogger() + + +@dataclass +class PreprocessedConditions: + clip_f: torch.Tensor + sync_f: torch.Tensor + text_f: torch.Tensor + clip_f_c: torch.Tensor + text_f_c: torch.Tensor + + +# Partially from https://github.com/facebookresearch/DiT +class MMAudio(nn.Module): + + def __init__(self, + *, + latent_dim: int, + clip_dim: int, + sync_dim: int, + text_dim: int, + hidden_dim: int, + depth: int, + fused_depth: int, + num_heads: int, + mlp_ratio: float = 4.0, + latent_seq_len: int, + clip_seq_len: int, + sync_seq_len: int, + text_seq_len: int = 77, + latent_mean: Optional[torch.Tensor] = None, + latent_std: Optional[torch.Tensor] = None, + empty_string_feat: Optional[torch.Tensor] = None, + v2: bool = False) -> None: + super().__init__() + + self.v2 = v2 + self.latent_dim = latent_dim + self._latent_seq_len = latent_seq_len + self._clip_seq_len = clip_seq_len + self._sync_seq_len = sync_seq_len + self._text_seq_len = text_seq_len + self.hidden_dim = hidden_dim + self.num_heads = num_heads + + if v2: + self.audio_input_proj = nn.Sequential( + ChannelLastConv1d(latent_dim, hidden_dim, kernel_size=7, padding=3), + nn.SiLU(), + ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=7, padding=3), + ) + + self.clip_input_proj = nn.Sequential( + nn.Linear(clip_dim, hidden_dim), + nn.SiLU(), + ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1), + ) + + self.sync_input_proj = nn.Sequential( + ChannelLastConv1d(sync_dim, hidden_dim, kernel_size=7, padding=3), + nn.SiLU(), + ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1), + ) + + self.text_input_proj = nn.Sequential( + nn.Linear(text_dim, hidden_dim), + nn.SiLU(), + MLP(hidden_dim, hidden_dim * 4), + ) + else: + self.audio_input_proj = nn.Sequential( + ChannelLastConv1d(latent_dim, hidden_dim, kernel_size=7, padding=3), + nn.SELU(), + ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=7, padding=3), + ) + + self.clip_input_proj = nn.Sequential( + nn.Linear(clip_dim, hidden_dim), + ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1), + ) + + self.sync_input_proj = nn.Sequential( + ChannelLastConv1d(sync_dim, hidden_dim, kernel_size=7, padding=3), + nn.SELU(), + ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1), + ) + + self.text_input_proj = nn.Sequential( + nn.Linear(text_dim, hidden_dim), + MLP(hidden_dim, hidden_dim * 4), + ) + + self.clip_cond_proj = nn.Linear(hidden_dim, hidden_dim) + self.text_cond_proj = nn.Linear(hidden_dim, hidden_dim) + self.global_cond_mlp = MLP(hidden_dim, hidden_dim * 4) + # each synchformer output segment has 8 feature frames + self.sync_pos_emb = nn.Parameter(torch.zeros((1, 1, 8, sync_dim))) + + self.final_layer = FinalBlock(hidden_dim, latent_dim) + + if v2: + self.t_embed = TimestepEmbedder(hidden_dim, + frequency_embedding_size=hidden_dim, + max_period=1) + else: + self.t_embed = TimestepEmbedder(hidden_dim, + frequency_embedding_size=256, + max_period=10000) + self.joint_blocks = nn.ModuleList([ + JointBlock(hidden_dim, + num_heads, + mlp_ratio=mlp_ratio, + pre_only=(i == depth - fused_depth - 1)) for i in range(depth - fused_depth) + ]) + + self.fused_blocks = nn.ModuleList([ + MMDitSingleBlock(hidden_dim, num_heads, mlp_ratio=mlp_ratio, kernel_size=3, padding=1) + for i in range(fused_depth) + ]) + + if latent_mean is None: + # these values are not meant to be used + # if you don't provide mean/std here, we should load them later from a checkpoint + assert latent_std is None + latent_mean = torch.ones(latent_dim).view(1, 1, -1).fill_(float('nan')) + latent_std = torch.ones(latent_dim).view(1, 1, -1).fill_(float('nan')) + else: + assert latent_std is not None + assert latent_mean.numel() == latent_dim, f'{latent_mean.numel()=} != {latent_dim=}' + if empty_string_feat is None: + empty_string_feat = torch.zeros((text_seq_len, text_dim)) + self.latent_mean = nn.Parameter(latent_mean.view(1, 1, -1), requires_grad=False) + self.latent_std = nn.Parameter(latent_std.view(1, 1, -1), requires_grad=False) + + self.empty_string_feat = nn.Parameter(empty_string_feat, requires_grad=False) + self.empty_clip_feat = nn.Parameter(torch.zeros(1, clip_dim), requires_grad=True) + self.empty_sync_feat = nn.Parameter(torch.zeros(1, sync_dim), requires_grad=True) + + self.initialize_weights() + self.initialize_rotations() + + def initialize_rotations(self): + base_freq = 1.0 + latent_rot = compute_rope_rotations(self._latent_seq_len, + self.hidden_dim // self.num_heads, + 10000, + freq_scaling=base_freq, + device=self.device) + clip_rot = compute_rope_rotations(self._clip_seq_len, + self.hidden_dim // self.num_heads, + 10000, + freq_scaling=base_freq * self._latent_seq_len / + self._clip_seq_len, + device=self.device) + + self.latent_rot = nn.Buffer(latent_rot, persistent=False) + self.clip_rot = nn.Buffer(clip_rot, persistent=False) + + def update_seq_lengths(self, latent_seq_len: int, clip_seq_len: int, sync_seq_len: int) -> None: + self._latent_seq_len = latent_seq_len + self._clip_seq_len = clip_seq_len + self._sync_seq_len = sync_seq_len + self.initialize_rotations() + + def initialize_weights(self): + + def _basic_init(module): + if isinstance(module, nn.Linear): + torch.nn.init.xavier_uniform_(module.weight) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + + self.apply(_basic_init) + + # Initialize timestep embedding MLP: + nn.init.normal_(self.t_embed.mlp[0].weight, std=0.02) + nn.init.normal_(self.t_embed.mlp[2].weight, std=0.02) + + # Zero-out adaLN modulation layers in DiT blocks: + for block in self.joint_blocks: + nn.init.constant_(block.latent_block.adaLN_modulation[-1].weight, 0) + nn.init.constant_(block.latent_block.adaLN_modulation[-1].bias, 0) + nn.init.constant_(block.clip_block.adaLN_modulation[-1].weight, 0) + nn.init.constant_(block.clip_block.adaLN_modulation[-1].bias, 0) + nn.init.constant_(block.text_block.adaLN_modulation[-1].weight, 0) + nn.init.constant_(block.text_block.adaLN_modulation[-1].bias, 0) + for block in self.fused_blocks: + nn.init.constant_(block.adaLN_modulation[-1].weight, 0) + nn.init.constant_(block.adaLN_modulation[-1].bias, 0) + + # Zero-out output layers: + nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0) + nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0) + nn.init.constant_(self.final_layer.conv.weight, 0) + nn.init.constant_(self.final_layer.conv.bias, 0) + + # empty string feat shall be initialized by a CLIP encoder + nn.init.constant_(self.sync_pos_emb, 0) + nn.init.constant_(self.empty_clip_feat, 0) + nn.init.constant_(self.empty_sync_feat, 0) + + def normalize(self, x: torch.Tensor) -> torch.Tensor: + # return (x - self.latent_mean) / self.latent_std + return x.sub_(self.latent_mean).div_(self.latent_std) + + def unnormalize(self, x: torch.Tensor) -> torch.Tensor: + # return x * self.latent_std + self.latent_mean + return x.mul_(self.latent_std).add_(self.latent_mean) + + def preprocess_conditions(self, clip_f: torch.Tensor, sync_f: torch.Tensor, + text_f: torch.Tensor) -> PreprocessedConditions: + """ + cache computations that do not depend on the latent/time step + i.e., the features are reused over steps during inference + """ + assert clip_f.shape[1] == self._clip_seq_len, f'{clip_f.shape=} {self._clip_seq_len=}' + assert sync_f.shape[1] == self._sync_seq_len, f'{sync_f.shape=} {self._sync_seq_len=}' + assert text_f.shape[1] == self._text_seq_len, f'{text_f.shape=} {self._text_seq_len=}' + + bs = clip_f.shape[0] + + # B * num_segments (24) * 8 * 768 + num_sync_segments = self._sync_seq_len // 8 + sync_f = sync_f.view(bs, num_sync_segments, 8, -1) + self.sync_pos_emb + sync_f = sync_f.flatten(1, 2) # (B, VN, D) + + # extend vf to match x + clip_f = self.clip_input_proj(clip_f) # (B, VN, D) + sync_f = self.sync_input_proj(sync_f) # (B, VN, D) + text_f = self.text_input_proj(text_f) # (B, VN, D) + + # upsample the sync features to match the audio + sync_f = sync_f.transpose(1, 2) # (B, D, VN) + sync_f = F.interpolate(sync_f, size=self._latent_seq_len, mode='nearest-exact') + sync_f = sync_f.transpose(1, 2) # (B, N, D) + + # get conditional features from the clip side + clip_f_c = self.clip_cond_proj(clip_f.mean(dim=1)) # (B, D) + text_f_c = self.text_cond_proj(text_f.mean(dim=1)) # (B, D) + + return PreprocessedConditions(clip_f=clip_f, + sync_f=sync_f, + text_f=text_f, + clip_f_c=clip_f_c, + text_f_c=text_f_c) + + def predict_flow(self, latent: torch.Tensor, t: torch.Tensor, + conditions: PreprocessedConditions) -> torch.Tensor: + """ + for non-cacheable computations + """ + assert latent.shape[1] == self._latent_seq_len, f'{latent.shape=} {self._latent_seq_len=}' + + clip_f = conditions.clip_f + sync_f = conditions.sync_f + text_f = conditions.text_f + clip_f_c = conditions.clip_f_c + text_f_c = conditions.text_f_c + + latent = self.audio_input_proj(latent) # (B, N, D) + global_c = self.global_cond_mlp(clip_f_c + text_f_c) # (B, D) + + global_c = self.t_embed(t).unsqueeze(1) + global_c.unsqueeze(1) # (B, D) + extended_c = global_c + sync_f + + for block in self.joint_blocks: + latent, clip_f, text_f = block(latent, clip_f, text_f, global_c, extended_c, + self.latent_rot, self.clip_rot) # (B, N, D) + + for block in self.fused_blocks: + latent = block(latent, extended_c, self.latent_rot) + + flow = self.final_layer(latent, global_c) # (B, N, out_dim), remove t + return flow + + def forward(self, latent: torch.Tensor, clip_f: torch.Tensor, sync_f: torch.Tensor, + text_f: torch.Tensor, t: torch.Tensor) -> torch.Tensor: + """ + latent: (B, N, C) + vf: (B, T, C_V) + t: (B,) + """ + conditions = self.preprocess_conditions(clip_f, sync_f, text_f) + flow = self.predict_flow(latent, t, conditions) + return flow + + def get_empty_string_sequence(self, bs: int) -> torch.Tensor: + return self.empty_string_feat.unsqueeze(0).expand(bs, -1, -1) + + def get_empty_clip_sequence(self, bs: int) -> torch.Tensor: + return self.empty_clip_feat.unsqueeze(0).expand(bs, self._clip_seq_len, -1) + + def get_empty_sync_sequence(self, bs: int) -> torch.Tensor: + return self.empty_sync_feat.unsqueeze(0).expand(bs, self._sync_seq_len, -1) + + def get_empty_conditions( + self, + bs: int, + *, + negative_text_features: Optional[torch.Tensor] = None) -> PreprocessedConditions: + if negative_text_features is not None: + empty_text = negative_text_features + else: + empty_text = self.get_empty_string_sequence(1) + + empty_clip = self.get_empty_clip_sequence(1) + empty_sync = self.get_empty_sync_sequence(1) + conditions = self.preprocess_conditions(empty_clip, empty_sync, empty_text) + conditions.clip_f = conditions.clip_f.expand(bs, -1, -1) + conditions.sync_f = conditions.sync_f.expand(bs, -1, -1) + conditions.clip_f_c = conditions.clip_f_c.expand(bs, -1) + if negative_text_features is None: + conditions.text_f = conditions.text_f.expand(bs, -1, -1) + conditions.text_f_c = conditions.text_f_c.expand(bs, -1) + + return conditions + + def ode_wrapper(self, t: torch.Tensor, latent: torch.Tensor, conditions: PreprocessedConditions, + empty_conditions: PreprocessedConditions, cfg_strength: float) -> torch.Tensor: + t = t * torch.ones(len(latent), device=latent.device, dtype=latent.dtype) + + if cfg_strength < 1.0: + return self.predict_flow(latent, t, conditions) + else: + return (cfg_strength * self.predict_flow(latent, t, conditions) + + (1 - cfg_strength) * self.predict_flow(latent, t, empty_conditions)) + + def load_weights(self, src_dict) -> None: + if 't_embed.freqs' in src_dict: + del src_dict['t_embed.freqs'] + if 'latent_rot' in src_dict: + del src_dict['latent_rot'] + if 'clip_rot' in src_dict: + del src_dict['clip_rot'] + + self.load_state_dict(src_dict, strict=True) + + @property + def device(self) -> torch.device: + return self.latent_mean.device + + @property + def latent_seq_len(self) -> int: + return self._latent_seq_len + + @property + def clip_seq_len(self) -> int: + return self._clip_seq_len + + @property + def sync_seq_len(self) -> int: + return self._sync_seq_len + + +def small_16k(**kwargs) -> MMAudio: + num_heads = 7 + return MMAudio(latent_dim=20, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=12, + fused_depth=8, + num_heads=num_heads, + latent_seq_len=250, + clip_seq_len=64, + sync_seq_len=192, + **kwargs) + + +def small_44k(**kwargs) -> MMAudio: + num_heads = 7 + return MMAudio(latent_dim=40, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=12, + fused_depth=8, + num_heads=num_heads, + latent_seq_len=345, + clip_seq_len=64, + sync_seq_len=192, + **kwargs) + + +def medium_44k(**kwargs) -> MMAudio: + num_heads = 14 + return MMAudio(latent_dim=40, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=12, + fused_depth=8, + num_heads=num_heads, + latent_seq_len=345, + clip_seq_len=64, + sync_seq_len=192, + **kwargs) + + +def large_44k(**kwargs) -> MMAudio: + num_heads = 14 + return MMAudio(latent_dim=40, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=21, + fused_depth=14, + num_heads=num_heads, + latent_seq_len=345, + clip_seq_len=64, + sync_seq_len=192, + **kwargs) + + +def large_44k_v2(**kwargs) -> MMAudio: + num_heads = 14 + return MMAudio(latent_dim=40, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=21, + fused_depth=14, + num_heads=num_heads, + latent_seq_len=345, + clip_seq_len=64, + sync_seq_len=192, + v2=True, + **kwargs) + + +def get_my_mmaudio(name: str, **kwargs) -> MMAudio: + if name == 'small_16k': + return small_16k(**kwargs) + if name == 'small_44k': + return small_44k(**kwargs) + if name == 'medium_44k': + return medium_44k(**kwargs) + if name == 'large_44k': + return large_44k(**kwargs) + if name == 'large_44k_v2': + return large_44k_v2(**kwargs) + + raise ValueError(f'Unknown model name: {name}') + + +if __name__ == '__main__': + network = get_my_mmaudio('small_16k') + + # print the number of parameters in terms of millions + num_params = sum(p.numel() for p in network.parameters()) / 1e6 + print(f'Number of parameters: {num_params:.2f}M') diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/sequence_config.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/sequence_config.py new file mode 100644 index 0000000000000000000000000000000000000000..14269014dc401b4751d172466813a935fddda6c1 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/sequence_config.py @@ -0,0 +1,58 @@ +import dataclasses +import math + + +@dataclasses.dataclass +class SequenceConfig: + # general + duration: float + + # audio + sampling_rate: int + spectrogram_frame_rate: int + latent_downsample_rate: int = 2 + + # visual + clip_frame_rate: int = 8 + sync_frame_rate: int = 25 + sync_num_frames_per_segment: int = 16 + sync_step_size: int = 8 + sync_downsample_rate: int = 2 + + @property + def num_audio_frames(self) -> int: + # we need an integer number of latents + return self.latent_seq_len * self.spectrogram_frame_rate * self.latent_downsample_rate + + @property + def latent_seq_len(self) -> int: + return int( + math.ceil(self.duration * self.sampling_rate / self.spectrogram_frame_rate / + self.latent_downsample_rate)) + + @property + def clip_seq_len(self) -> int: + return int(self.duration * self.clip_frame_rate) + + @property + def sync_seq_len(self) -> int: + num_frames = self.duration * self.sync_frame_rate + num_segments = (num_frames - self.sync_num_frames_per_segment) // self.sync_step_size + 1 + return int(num_segments * self.sync_num_frames_per_segment / self.sync_downsample_rate) + + +CONFIG_16K = SequenceConfig(duration=8.0, sampling_rate=16000, spectrogram_frame_rate=256) +CONFIG_44K = SequenceConfig(duration=8.0, sampling_rate=44100, spectrogram_frame_rate=512) + +if __name__ == '__main__': + assert CONFIG_16K.latent_seq_len == 250 + assert CONFIG_16K.clip_seq_len == 64 + assert CONFIG_16K.sync_seq_len == 192 + assert CONFIG_16K.num_audio_frames == 128000 + + assert CONFIG_44K.latent_seq_len == 345 + assert CONFIG_44K.clip_seq_len == 64 + assert CONFIG_44K.sync_seq_len == 192 + assert CONFIG_44K.num_audio_frames == 353280 + + print('Passed') diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/transformer_layers.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/transformer_layers.py new file mode 100644 index 0000000000000000000000000000000000000000..c3ac849a37b52329406d86b12f2f5c9ddd0e9823 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/transformer_layers.py @@ -0,0 +1,203 @@ +from typing import Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange +from einops.layers.torch import Rearrange +from torch.nn.attention import SDPBackend, sdpa_kernel + +from ..ext.rotary_embeddings import apply_rope +from .low_level import MLP, ChannelLastConv1d, ConvMLP + + +def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor): + return x * (1 + scale) + shift + + +def attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor): + # training will crash without these contiguous calls and the CUDNN limitation + # I believe this is related to https://github.com/pytorch/pytorch/issues/133974 + # unresolved at the time of writing + q = q.contiguous() + k = k.contiguous() + v = v.contiguous() + out = F.scaled_dot_product_attention(q, k, v) + out = rearrange(out, 'b h n d -> b n (h d)').contiguous() + return out + + +class SelfAttention(nn.Module): + + def __init__(self, dim: int, nheads: int): + super().__init__() + self.dim = dim + self.nheads = nheads + + self.qkv = nn.Linear(dim, dim * 3, bias=True) + self.q_norm = nn.RMSNorm(dim // nheads) + self.k_norm = nn.RMSNorm(dim // nheads) + + self.split_into_heads = Rearrange('b n (h d j) -> b h n d j', + h=nheads, + d=dim // nheads, + j=3) + + def pre_attention( + self, x: torch.Tensor, + rot: Optional[torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + # x: batch_size * n_tokens * n_channels + qkv = self.qkv(x) + q, k, v = self.split_into_heads(qkv).chunk(3, dim=-1) + q = q.squeeze(-1) + k = k.squeeze(-1) + v = v.squeeze(-1) + q = self.q_norm(q) + k = self.k_norm(k) + + if rot is not None: + q = apply_rope(q, rot) + k = apply_rope(k, rot) + + return q, k, v + + def forward( + self, + x: torch.Tensor, # batch_size * n_tokens * n_channels + ) -> torch.Tensor: + q, v, k = self.pre_attention(x) + out = attention(q, k, v) + return out + + +class MMDitSingleBlock(nn.Module): + + def __init__(self, + dim: int, + nhead: int, + mlp_ratio: float = 4.0, + pre_only: bool = False, + kernel_size: int = 7, + padding: int = 3): + super().__init__() + self.norm1 = nn.LayerNorm(dim, elementwise_affine=False) + self.attn = SelfAttention(dim, nhead) + + self.pre_only = pre_only + if pre_only: + self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 2 * dim, bias=True)) + else: + if kernel_size == 1: + self.linear1 = nn.Linear(dim, dim) + else: + self.linear1 = ChannelLastConv1d(dim, dim, kernel_size=kernel_size, padding=padding) + self.norm2 = nn.LayerNorm(dim, elementwise_affine=False) + + if kernel_size == 1: + self.ffn = MLP(dim, int(dim * mlp_ratio)) + else: + self.ffn = ConvMLP(dim, + int(dim * mlp_ratio), + kernel_size=kernel_size, + padding=padding) + + self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 6 * dim, bias=True)) + + def pre_attention(self, x: torch.Tensor, c: torch.Tensor, rot: Optional[torch.Tensor]): + # x: BS * N * D + # cond: BS * D + modulation = self.adaLN_modulation(c) + if self.pre_only: + (shift_msa, scale_msa) = modulation.chunk(2, dim=-1) + gate_msa = shift_mlp = scale_mlp = gate_mlp = None + else: + (shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, + gate_mlp) = modulation.chunk(6, dim=-1) + + x = modulate(self.norm1(x), shift_msa, scale_msa) + q, k, v = self.attn.pre_attention(x, rot) + return (q, k, v), (gate_msa, shift_mlp, scale_mlp, gate_mlp) + + def post_attention(self, x: torch.Tensor, attn_out: torch.Tensor, c: tuple[torch.Tensor]): + if self.pre_only: + return x + + (gate_msa, shift_mlp, scale_mlp, gate_mlp) = c + x = x + self.linear1(attn_out) * gate_msa + r = modulate(self.norm2(x), shift_mlp, scale_mlp) + x = x + self.ffn(r) * gate_mlp + + return x + + def forward(self, x: torch.Tensor, cond: torch.Tensor, + rot: Optional[torch.Tensor]) -> torch.Tensor: + # x: BS * N * D + # cond: BS * D + x_qkv, x_conditions = self.pre_attention(x, cond, rot) + attn_out = attention(*x_qkv) + x = self.post_attention(x, attn_out, x_conditions) + + return x + + +class JointBlock(nn.Module): + + def __init__(self, dim: int, nhead: int, mlp_ratio: float = 4.0, pre_only: bool = False): + super().__init__() + self.pre_only = pre_only + self.latent_block = MMDitSingleBlock(dim, + nhead, + mlp_ratio, + pre_only=False, + kernel_size=3, + padding=1) + self.clip_block = MMDitSingleBlock(dim, + nhead, + mlp_ratio, + pre_only=pre_only, + kernel_size=3, + padding=1) + self.text_block = MMDitSingleBlock(dim, nhead, mlp_ratio, pre_only=pre_only, kernel_size=1) + + def forward(self, latent: torch.Tensor, clip_f: torch.Tensor, text_f: torch.Tensor, + global_c: torch.Tensor, extended_c: torch.Tensor, latent_rot: torch.Tensor, + clip_rot: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + # latent: BS * N1 * D + # clip_f: BS * N2 * D + # c: BS * (1/N) * D + x_qkv, x_mod = self.latent_block.pre_attention(latent, extended_c, latent_rot) + c_qkv, c_mod = self.clip_block.pre_attention(clip_f, global_c, clip_rot) + t_qkv, t_mod = self.text_block.pre_attention(text_f, global_c, rot=None) + + latent_len = latent.shape[1] + clip_len = clip_f.shape[1] + text_len = text_f.shape[1] + + joint_qkv = [torch.cat([x_qkv[i], c_qkv[i], t_qkv[i]], dim=2) for i in range(3)] + + attn_out = attention(*joint_qkv) + x_attn_out = attn_out[:, :latent_len] + c_attn_out = attn_out[:, latent_len:latent_len + clip_len] + t_attn_out = attn_out[:, latent_len + clip_len:] + + latent = self.latent_block.post_attention(latent, x_attn_out, x_mod) + if not self.pre_only: + clip_f = self.clip_block.post_attention(clip_f, c_attn_out, c_mod) + text_f = self.text_block.post_attention(text_f, t_attn_out, t_mod) + + return latent, clip_f, text_f + + +class FinalBlock(nn.Module): + + def __init__(self, dim, out_dim): + super().__init__() + self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 2 * dim, bias=True)) + self.norm = nn.LayerNorm(dim, elementwise_affine=False) + self.conv = ChannelLastConv1d(dim, out_dim, kernel_size=7, padding=3) + + def forward(self, latent, c): + shift, scale = self.adaLN_modulation(c).chunk(2, dim=-1) + latent = modulate(self.norm(latent), shift, scale) + latent = self.conv(latent) + return latent diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/distributions.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/distributions.py new file mode 100644 index 0000000000000000000000000000000000000000..1d526a5b0b3dd2ae556d806a3397e1cf43c07fb9 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/distributions.py @@ -0,0 +1,46 @@ +from typing import Optional + +import numpy as np +import torch + + +class DiagonalGaussianDistribution: + + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self, rng: Optional[torch.Generator] = None): + # x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) + + r = torch.empty_like(self.mean).normal_(generator=rng) + x = self.mean + self.std * r + + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + + return 0.5 * torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar + else: + return 0.5 * (torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar) + + def nll(self, sample, dims=[1, 2, 3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum(logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/features_utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/features_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2157d14bc2fd9e7e72397c620b39555aca92d7ba --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/features_utils.py @@ -0,0 +1,155 @@ +from typing import Literal, Optional + +import open_clip +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange +from torchvision.transforms import Normalize + + +from ...ext.mel_converter import MelConverter +from ...model.utils.distributions import DiagonalGaussianDistribution +from ...ext.bigvgan import BigVGAN + + +def patch_clip(clip_model): + # a hack to make it output last hidden states + # https://github.com/mlfoundations/open_clip/blob/fc5a37b72d705f760ebbc7915b84729816ed471f/src/open_clip/model.py#L269 + def new_encode_text(self, text, normalize: bool = False): + cast_dtype = self.transformer.get_cast_dtype() + + x = self.token_embedding(text).to(cast_dtype) # [batch_size, n_ctx, d_model] + + x = x + self.positional_embedding.to(cast_dtype) + x = self.transformer(x, attn_mask=self.attn_mask) + x = self.ln_final(x) # [batch_size, n_ctx, transformer.width] + return F.normalize(x, dim=-1) if normalize else x + + clip_model.encode_text = new_encode_text.__get__(clip_model) + return clip_model + + +class FeaturesUtils(nn.Module): + + def __init__( + self, + *, + vae: None, + synchformer: None, + clip_model: None, + enable_conditions: bool = True, + ): + super().__init__() + + if enable_conditions: + #self.clip_model = create_model_from_pretrained('hf-hub:apple/DFN5B-CLIP-ViT-H-14-384', + # return_transform=False) + self.clip_model = clip_model + self.clip_preprocess = Normalize(mean=[0.48145466, 0.4578275, 0.40821073], + std=[0.26862954, 0.26130258, 0.27577711]) + self.clip_model = patch_clip(self.clip_model) + + self.synchformer = synchformer + + self.tokenizer = open_clip.get_tokenizer('ViT-H-14-378-quickgelu') # same as 'ViT-H-14' + else: + self.clip_model = None + self.synchformer = None + self.tokenizer = None + + + self.tod = vae + self.mel_converter = MelConverter() + + def compile(self): + if self.clip_model is not None: + self.encode_video_with_clip = torch.compile(self.encode_video_with_clip) + self.clip_model.encode_image = torch.compile(self.clip_model.encode_image) + self.clip_model.encode_text = torch.compile(self.clip_model.encode_text) + if self.synchformer is not None: + self.synchformer = torch.compile(self.synchformer) + self.tod.encode = torch.compile(self.tod.encode) + self.decode = torch.compile(self.decode) + self.vocode = torch.compile(self.vocode) + + def train(self, mode: bool) -> None: + return super().train(False) + + @torch.inference_mode() + def encode_video_with_clip(self, x: torch.Tensor, batch_size: int = -1) -> torch.Tensor: + assert self.clip_model is not None, 'CLIP is not loaded' + # x: (B, T, C, H, W) H/W: 384 + b, t, c, h, w = x.shape + assert c == 3 and h == 384 and w == 384 + x = self.clip_preprocess(x) + x = rearrange(x, 'b t c h w -> (b t) c h w') + outputs = [] + if batch_size < 0: + batch_size = b * t + for i in range(0, b * t, batch_size): + outputs.append(self.clip_model.encode_image(x[i:i + batch_size], normalize=True)) + x = torch.cat(outputs, dim=0) + # x = self.clip_model.encode_image(x, normalize=True) + x = rearrange(x, '(b t) d -> b t d', b=b) + return x + + @torch.inference_mode() + def encode_video_with_sync(self, x: torch.Tensor, batch_size: int = -1) -> torch.Tensor: + assert self.synchformer is not None, 'Synchformer is not loaded' + # x: (B, T, C, H, W) H/W: 384 + + b, t, c, h, w = x.shape + assert c == 3 and h == 224 and w == 224 + + # partition the video + segment_size = 16 + step_size = 8 + num_segments = (t - segment_size) // step_size + 1 + segments = [] + for i in range(num_segments): + segments.append(x[:, i * step_size:i * step_size + segment_size]) + x = torch.stack(segments, dim=1) # (B, S, T, C, H, W) + + outputs = [] + if batch_size < 0: + batch_size = b + for i in range(0, b, batch_size): + outputs.append(self.synchformer(x[i:i + batch_size])) + x = torch.cat(outputs, dim=0).flatten(start_dim=1, end_dim=2) + return x + + @torch.inference_mode() + def encode_text(self, text: list[str]) -> torch.Tensor: + assert self.clip_model is not None, 'CLIP is not loaded' + assert self.tokenizer is not None, 'Tokenizer is not loaded' + # x: (B, L) + tokens = self.tokenizer(text).to(self.device) + return self.clip_model.encode_text(tokens, normalize=True) + + @torch.inference_mode() + def encode_audio(self, x) -> DiagonalGaussianDistribution: + assert self.tod is not None, 'VAE is not loaded' + # x: (B * L) + mel = self.mel_converter(x) + dist = self.tod.encode(mel) + + return dist + + @torch.inference_mode() + def vocode(self, mel: torch.Tensor) -> torch.Tensor: + assert self.tod is not None, 'VAE is not loaded' + return self.tod.vocode(mel) + + @torch.inference_mode() + def decode(self, z: torch.Tensor) -> torch.Tensor: + assert self.tod is not None, 'VAE is not loaded' + return self.tod.decode(z.transpose(1, 2)) + + @property + def device(self): + return next(self.parameters()).device + + @property + def dtype(self): + return next(self.parameters()).dtype diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/parameter_groups.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/parameter_groups.py new file mode 100644 index 0000000000000000000000000000000000000000..89c3993083f470dfc6b18a5c90f908ea37bde12b --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/parameter_groups.py @@ -0,0 +1,72 @@ +import logging + +log = logging.getLogger() + + +def get_parameter_groups(model, cfg, print_log=False): + """ + Assign different weight decays and learning rates to different parameters. + Returns a parameter group which can be passed to the optimizer. + """ + weight_decay = cfg.weight_decay + # embed_weight_decay = cfg.embed_weight_decay + # backbone_lr_ratio = cfg.backbone_lr_ratio + base_lr = cfg.learning_rate + + backbone_params = [] + embed_params = [] + other_params = [] + + # embedding_names = ['summary_pos', 'query_init', 'query_emb', 'obj_pe'] + # embedding_names = [e + '.weight' for e in embedding_names] + + # inspired by detectron2 + memo = set() + for name, param in model.named_parameters(): + if not param.requires_grad: + continue + # Avoid duplicating parameters + if param in memo: + continue + memo.add(param) + + if name.startswith('module'): + name = name[7:] + + inserted = False + # if name.startswith('pixel_encoder.'): + # backbone_params.append(param) + # inserted = True + # if print_log: + # log.info(f'{name} counted as a backbone parameter.') + # else: + # for e in embedding_names: + # if name.endswith(e): + # embed_params.append(param) + # inserted = True + # if print_log: + # log.info(f'{name} counted as an embedding parameter.') + # break + + # if not inserted: + other_params.append(param) + + parameter_groups = [ + # { + # 'params': backbone_params, + # 'lr': base_lr * backbone_lr_ratio, + # 'weight_decay': weight_decay + # }, + # { + # 'params': embed_params, + # 'lr': base_lr, + # 'weight_decay': embed_weight_decay + # }, + { + 'params': other_params, + 'lr': base_lr, + 'weight_decay': weight_decay + }, + ] + + return parameter_groups diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/sample_utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/sample_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d44cf278e0b464bc6ac7e240fcab4a23895caa2f --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/model/utils/sample_utils.py @@ -0,0 +1,12 @@ +from typing import Optional + +import torch + + +def log_normal_sample(x: torch.Tensor, + generator: Optional[torch.Generator] = None, + m: float = 0.0, + s: float = 1.0) -> torch.Tensor: + bs = x.shape[0] + s = torch.randn(bs, device=x.device, generator=generator) * s + m + return torch.sigmoid(s) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/__init__.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/dist_utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/dist_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..28684263c8b0ede6f6b600f1087d767dbe972281 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/dist_utils.py @@ -0,0 +1,17 @@ +import os +from logging import Logger + +from .utils.logger import TensorboardLogger + +local_rank = int(os.environ['LOCAL_RANK']) if 'LOCAL_RANK' in os.environ else 0 +world_size = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1 + + +def info_if_rank_zero(logger: Logger, msg: str): + if local_rank == 0: + logger.info(msg) + + +def string_if_rank_zero(logger: TensorboardLogger, tag: str, msg: str): + if local_rank == 0: + logger.log_string(tag, msg) diff --git a/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/download_utils.py b/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/download_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..09c48f5acd90e277a7b80d5ddddb37566948361c --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/mmaudio/utils/download_utils.py @@ -0,0 +1,84 @@ +import hashlib +import logging +from pathlib import Path + +import requests +from tqdm import tqdm + +log = logging.getLogger() + +links = [ + { + 'name': 'mmaudio_small_16k.pth', + 'url': 'https://databank.illinois.edu/datafiles/k6jve/download', + 'md5': 'af93cde404179f58e3919ac085b8033b', + }, + { + 'name': 'mmaudio_small_44k.pth', + 'url': 'https://databank.illinois.edu/datafiles/864ya/download', + 'md5': 'babd74c884783d13701ea2820a5f5b6d', + }, + { + 'name': 'mmaudio_medium_44k.pth', + 'url': 'https://databank.illinois.edu/datafiles/pa94t/download', + 'md5': '5a56b6665e45a1e65ada534defa903d0', + }, + { + 'name': 'mmaudio_large_44k.pth', + 'url': 'https://databank.illinois.edu/datafiles/4jx76/download', + 'md5': 'fed96c325a6785b85ce75ae1aafd2673' + }, + { + 'name': 'mmaudio_large_44k_v2.pth', + 'url': 'https://databank.illinois.edu/datafiles/16j46/download', + 'md5': '01ad4464f049b2d7efdaa4c1a59b8dfe' + }, + { + 'name': 'v1-16.pth', + 'url': 'https://github.com/hkchengrex/MMAudio/releases/download/v0.1/v1-16.pth', + 'md5': '69f56803f59a549a1a507c93859fd4d7' + }, + { + 'name': 'best_netG.pt', + 'url': 'https://github.com/hkchengrex/MMAudio/releases/download/v0.1/best_netG.pt', + 'md5': 'eeaf372a38a9c31c362120aba2dde292' + }, + { + 'name': 'v1-44.pth', + 'url': 'https://github.com/hkchengrex/MMAudio/releases/download/v0.1/v1-44.pth', + 'md5': 'fab020275fa44c6589820ce025191600' + }, + { + 'name': 'synchformer_state_dict.pth', + 'url': + 'https://github.com/hkchengrex/MMAudio/releases/download/v0.1/synchformer_state_dict.pth', + 'md5': '5b2f5594b0730f70e41e549b7c94390c' + }, +] + + +def download_model_if_needed(model_path: Path): + base_name = model_path.name + + for link in links: + if link['name'] == base_name: + target_link = link + break + else: + raise ValueError(f'No link found for {base_name}') + + model_path.parent.mkdir(parents=True, exist_ok=True) + if not model_path.exists() or hashlib.md5(open(model_path, + 'rb').read()).hexdigest() != target_link['md5']: + log.info(f'Downloading {base_name} to {model_path}...') + r = requests.get(target_link['url'], stream=True) + total_size = int(r.headers.get('content-length', 0)) + block_size = 1024 + t = tqdm(total=total_size, unit='iB', unit_scale=True) + with open(model_path, 'wb') as f: + for data in r.iter_content(block_size): + t.update(len(data)) + f.write(data) + t.close() + if total_size != 0 and t.n != total_size: + raise RuntimeError('Error while downloading %s' % base_name) diff --git a/custom_nodes/ComfyUI-MMAudio/nodes.py b/custom_nodes/ComfyUI-MMAudio/nodes.py new file mode 100644 index 0000000000000000000000000000000000000000..5e92dfe1c0025f3426185f988f89e88ccd553d71 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/nodes.py @@ -0,0 +1,379 @@ +import os +import torch +import json +from torchvision.transforms import v2 +from accelerate import init_empty_weights +from accelerate.utils import set_module_tensor_to_device + +import folder_paths +import comfy.model_management as mm +from comfy.utils import load_torch_file + +script_directory = os.path.dirname(os.path.abspath(__file__)) + +if not "mmaudio" in folder_paths.folder_names_and_paths: + folder_paths.add_model_folder_path("mmaudio", os.path.join(folder_paths.models_dir, "mmaudio")) + +from .mmaudio.eval_utils import generate +from .mmaudio.model.flow_matching import FlowMatching +from .mmaudio.model.networks import MMAudio +from .mmaudio.model.utils.features_utils import FeaturesUtils +from .mmaudio.model.sequence_config import (CONFIG_16K, CONFIG_44K, SequenceConfig) +from .mmaudio.ext.bigvgan_v2.bigvgan import BigVGAN as BigVGANv2 +from .mmaudio.ext.synchformer import Synchformer +from .mmaudio.ext.autoencoder import AutoEncoderModule +from open_clip import CLIP + +import logging +logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') +log = logging.getLogger(__name__) + +def process_video_tensor(video_tensor: torch.Tensor, duration_sec: float) -> tuple[torch.Tensor, torch.Tensor, float]: + _CLIP_SIZE = 384 + _CLIP_FPS = 8.0 + + _SYNC_SIZE = 224 + _SYNC_FPS = 25.0 + + clip_transform = v2.Compose([ + v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC), + v2.ToPILImage(), + v2.ToTensor(), + v2.ConvertImageDtype(torch.float32), + ]) + + sync_transform = v2.Compose([ + v2.Resize(_SYNC_SIZE, interpolation=v2.InterpolationMode.BICUBIC), + v2.CenterCrop(_SYNC_SIZE), + v2.ToPILImage(), + v2.ToTensor(), + v2.ConvertImageDtype(torch.float32), + v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), + ]) + + # Assuming video_tensor is in the shape (frames, height, width, channels) + total_frames = video_tensor.shape[0] + clip_frames_count = int(_CLIP_FPS * duration_sec) + sync_frames_count = int(_SYNC_FPS * duration_sec) + + # Adjust duration if there are not enough frames + if total_frames < clip_frames_count: + log.warning(f'Clip video is too short: {total_frames / _CLIP_FPS:.2f} < {duration_sec:.2f}') + clip_frames_count = total_frames + duration_sec = total_frames / _CLIP_FPS + + if total_frames < sync_frames_count: + log.warning(f'Sync video is too short: {total_frames / _SYNC_FPS:.2f} < {duration_sec:.2f}, truncating to {total_frames / _SYNC_FPS:.2f} sec') + sync_frames_count = total_frames + duration_sec = total_frames / _SYNC_FPS + + clip_frames = video_tensor[:clip_frames_count] + sync_frames = video_tensor[:sync_frames_count] + + clip_frames = clip_frames.permute(0, 3, 1, 2) + sync_frames = sync_frames.permute(0, 3, 1, 2) + + clip_frames = torch.stack([clip_transform(frame) for frame in clip_frames]) + sync_frames = torch.stack([sync_transform(frame) for frame in sync_frames]) + + clip_length_sec = clip_frames.shape[0] / _CLIP_FPS + sync_length_sec = sync_frames.shape[0] / _SYNC_FPS + + # if clip_length_sec < duration_sec: + # log.warning(f'Clip video is too short: {clip_length_sec:.2f} < {duration_sec:.2f}') + # log.warning(f'Truncating to {clip_length_sec:.2f} sec') + # duration_sec = clip_length_sec + + # if sync_length_sec < duration_sec: + # log.warning(f'Sync video is too short: {sync_length_sec:.2f} < {duration_sec:.2f}') + # log.warning(f'Truncating to {sync_length_sec:.2f} sec') + # duration_sec = sync_length_sec + + clip_frames = clip_frames[:int(_CLIP_FPS * duration_sec)] + sync_frames = sync_frames[:int(_SYNC_FPS * duration_sec)] + + return clip_frames, sync_frames, duration_sec + +#region Model loading +class MMAudioModelLoader: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "mmaudio_model": (folder_paths.get_filename_list("mmaudio"), {"tooltip": "These models are loaded from the 'ComfyUI/models/mmaudio' -folder",}), + + "base_precision": (["fp16", "fp32", "bf16"], {"default": "fp16"}), + }, + } + + RETURN_TYPES = ("MMAUDIO_MODEL",) + RETURN_NAMES = ("mmaudio_model", ) + FUNCTION = "loadmodel" + CATEGORY = "MMAudio" + + def loadmodel(self, mmaudio_model, base_precision): + + + device = mm.get_torch_device() + offload_device = mm.unet_offload_device() + + mm.soft_empty_cache() + + base_dtype = {"fp8_e4m3fn": torch.float8_e4m3fn, "fp8_e4m3fn_fast": torch.float8_e4m3fn, "bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[base_precision] + + mmaudio_model_path = folder_paths.get_full_path_or_raise("mmaudio", mmaudio_model) + mmaudio_sd = load_torch_file(mmaudio_model_path, device=offload_device) + + with init_empty_weights(): + # small + if mmaudio_sd["audio_input_proj.0.bias"].shape[0] == 448: + num_heads = 7 + model = MMAudio( + latent_dim=40, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=12, + fused_depth=8, + num_heads=num_heads, + latent_seq_len=345, + clip_seq_len=64, + sync_seq_len=192 + ) + # large + elif mmaudio_sd["audio_input_proj.0.bias"].shape[0] == 896: + num_heads = 14 + model = MMAudio(latent_dim=40, + clip_dim=1024, + sync_dim=768, + text_dim=1024, + hidden_dim=64 * num_heads, + depth=21, + fused_depth=14, + num_heads=num_heads, + latent_seq_len=345, + clip_seq_len=64, + sync_seq_len=192, + v2=mmaudio_sd["t_embed.mlp.0.weight"].shape[1] == 896 + ) + model = model.eval() + for name, param in model.named_parameters(): + # Set tensor to device + set_module_tensor_to_device(model, name, device=offload_device, dtype=base_dtype, value=mmaudio_sd[name]) + del mmaudio_sd + log.info(f'Loaded MMAudio model weights from {mmaudio_model_path}') + if "44" in mmaudio_model: + model.seq_cfg = CONFIG_44K + elif "16" in mmaudio_model: + model.seq_cfg = CONFIG_16K + + return (model,) + +#region Features Utils +class MMAudioVoCoderLoader: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "vocoder_model": (folder_paths.get_filename_list("mmaudio"), {"tooltip": "These models are loaded from 'ComfyUI/models/mmaudio'"}), + }, + } + + RETURN_TYPES = ("VOCODER_MODEL",) + RETURN_NAMES = ("mmaudio_vocoder", ) + FUNCTION = "loadmodel" + CATEGORY = "MMAudio" + + def loadmodel(self, vocoder_model): + from .mmaudio.ext.bigvgan import BigVGAN + vocoder_model_path = folder_paths.get_full_path_or_raise("mmaudio", vocoder_model) + vocoder_model = BigVGAN.from_pretrained(vocoder_model_path).eval() + return (vocoder_model_path,) + +class MMAudioFeatureUtilsLoader: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "vae_model": (folder_paths.get_filename_list("mmaudio"), {"tooltip": "These models are loaded from 'ComfyUI/models/mmaudio'"}), + "synchformer_model": (folder_paths.get_filename_list("mmaudio"), {"tooltip": "These models are loaded from 'ComfyUI/models/mmaudio'"}), + "clip_model": (folder_paths.get_filename_list("mmaudio"), {"tooltip": "These models are loaded from 'ComfyUI/models/mmaudio'"}), + }, + "optional": { + "bigvgan_vocoder_model": ("VOCODER_MODEL", {"tooltip": "These models are loaded from 'ComfyUI/models/mmaudio'"}), + "mode": (["16k", "44k"], {"default": "44k"}), + "precision": (["fp16", "fp32", "bf16"], + {"default": "fp16"} + ), + } + } + + RETURN_TYPES = ("MMAUDIO_FEATUREUTILS",) + RETURN_NAMES = ("mmaudio_featureutils", ) + FUNCTION = "loadmodel" + CATEGORY = "MMAudio" + + def loadmodel(self, vae_model, precision, synchformer_model, clip_model, mode, bigvgan_vocoder_model=None): + + device = mm.get_torch_device() + offload_device = mm.unet_offload_device() + + dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[precision] + + #synchformer + synchformer_path = folder_paths.get_full_path_or_raise("mmaudio", synchformer_model) + synchformer_sd = load_torch_file(synchformer_path, device=offload_device) + with init_empty_weights(): + synchformer = Synchformer().eval() + + for name, param in synchformer.named_parameters(): + # Set tensor to device + set_module_tensor_to_device(synchformer, name, device=device, dtype=dtype, value=synchformer_sd[name]) + + #vae + download_path = folder_paths.get_folder_paths("mmaudio")[0] + + nvidia_bigvgan_vocoder_path = os.path.join(download_path, "nvidia", "bigvgan_v2_44khz_128band_512x") + if mode == "44k": + if not os.path.exists(nvidia_bigvgan_vocoder_path): + log.info(f"Downloading nvidia bigvgan vocoder model to: {nvidia_bigvgan_vocoder_path}") + from huggingface_hub import snapshot_download + + snapshot_download( + repo_id="nvidia/bigvgan_v2_44khz_128band_512x", + ignore_patterns=["*3m*",], + local_dir=nvidia_bigvgan_vocoder_path, + local_dir_use_symlinks=False, + ) + + bigvgan_vocoder = BigVGANv2.from_pretrained(nvidia_bigvgan_vocoder_path).eval().to(device=device, dtype=dtype) + else: + assert bigvgan_vocoder_model is not None, "bigvgan_vocoder_model must be provided for 16k mode" + bigvgan_vocoder = bigvgan_vocoder_model + + vae_path = folder_paths.get_full_path_or_raise("mmaudio", vae_model) + vae_sd = load_torch_file(vae_path, device=offload_device) + vae = AutoEncoderModule( + vae_state_dict=vae_sd, + bigvgan_vocoder=bigvgan_vocoder, + mode=mode + ) + vae = vae.eval().to(device=device, dtype=dtype) + + #clip + + clip_model_path = folder_paths.get_full_path_or_raise("mmaudio", clip_model) + clip_config_path = os.path.join(script_directory, "configs", "DFN5B-CLIP-ViT-H-14-384.json") + with open(clip_config_path) as f: + clip_config = json.load(f) + + with init_empty_weights(): + try: + clip_model = CLIP(**clip_config["model_cfg"]).eval() + except: + # for some open-clip versions + clip_config["model_cfg"]["nonscalar_logit_scale"] = True + clip_model = CLIP(**clip_config["model_cfg"]).eval() + + clip_sd = load_torch_file(os.path.join(clip_model_path), device=offload_device) + for name, param in clip_model.named_parameters(): + set_module_tensor_to_device(clip_model, name, device=device, dtype=dtype, value=clip_sd[name]) + clip_model.to(device=device, dtype=dtype) + + #clip_model = create_model_from_pretrained("hf-hub:apple/DFN5B-CLIP-ViT-H-14-384", return_transform=False) + + + feature_utils = FeaturesUtils(vae=vae, + synchformer=synchformer, + enable_conditions=True, + clip_model=clip_model) + return (feature_utils,) + +#region sampling +class MMAudioSampler: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "mmaudio_model": ("MMAUDIO_MODEL",), + "feature_utils": ("MMAUDIO_FEATUREUTILS",), + "duration": ("FLOAT", {"default": 8, "step": 0.01, "tooltip": "Duration of the audio in seconds"}), + "steps": ("INT", {"default": 25, "step": 1, "tooltip": "Number of steps to interpolate"}), + "cfg": ("FLOAT", {"default": 4.5, "step": 0.1, "tooltip": "Strength of the conditioning"}), + "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}), + "prompt": ("STRING", {"default": "", "multiline": True} ), + "negative_prompt": ("STRING", {"default": "", "multiline": True} ), + "mask_away_clip": ("BOOLEAN", {"default": False, "tooltip": "If true, the clip video will be masked away"}), + "force_offload": ("BOOLEAN", {"default": True, "tooltip": "If true, the model will be offloaded to the offload device"}), + }, + "optional": { + "images": ("IMAGE",), + }, + } + + RETURN_TYPES = ("AUDIO",) + RETURN_NAMES = ("audio", ) + FUNCTION = "sample" + CATEGORY = "MMAudio" + + def sample(self, mmaudio_model, seed, feature_utils, duration, steps, cfg, prompt, negative_prompt, mask_away_clip, force_offload, images=None): + device = mm.get_torch_device() + offload_device = mm.unet_offload_device() + rng = torch.Generator(device=device) + rng.manual_seed(seed) + + seq_cfg = mmaudio_model.seq_cfg + + if images is not None: + images = images.to(device=device) + clip_frames, sync_frames, duration = process_video_tensor(images, duration) + print("clip_frames", clip_frames.shape, "sync_frames", sync_frames.shape, "duration", duration) + if mask_away_clip: + clip_frames = None + else: + clip_frames = clip_frames.unsqueeze(0) + sync_frames = sync_frames.unsqueeze(0) + else: + clip_frames = None + sync_frames = None + + seq_cfg.duration = duration + mmaudio_model.update_seq_lengths(seq_cfg.latent_seq_len, seq_cfg.clip_seq_len, seq_cfg.sync_seq_len) + + scheduler = FlowMatching(min_sigma=0, inference_mode='euler', num_steps=steps) + feature_utils.to(device) + mmaudio_model.to(device) + audios = generate(clip_frames, + sync_frames, [prompt], + negative_text=[negative_prompt], + feature_utils=feature_utils, + net=mmaudio_model, + fm=scheduler, + rng=rng, + cfg_strength=cfg) + if force_offload: + mmaudio_model.to(offload_device) + feature_utils.to(offload_device) + mm.soft_empty_cache() + waveform = audios.float().cpu() + #torchaudio.save("test.wav", waveform, 44100) + audio = { + "waveform": waveform, + "sample_rate": 44100 + } + + return (audio,) + +NODE_CLASS_MAPPINGS = { + "MMAudioModelLoader": MMAudioModelLoader, + "MMAudioFeatureUtilsLoader": MMAudioFeatureUtilsLoader, + "MMAudioSampler": MMAudioSampler, + "MMAudioVoCoderLoader": MMAudioVoCoderLoader, +} +NODE_DISPLAY_NAME_MAPPINGS = { + "MMAudioModelLoader": "MMAudio ModelLoader", + "MMAudioFeatureUtilsLoader": "MMAudio FeatureUtilsLoader", + "MMAudioSampler": "MMAudio Sampler", + "MMAudioVoCoderLoader": "MMAudio VoCoderLoader", + } diff --git a/custom_nodes/ComfyUI-MMAudio/pyproject.toml b/custom_nodes/ComfyUI-MMAudio/pyproject.toml new file mode 100644 index 0000000000000000000000000000000000000000..b3d05bf665a5da4024a290669b7a21236d47e8d8 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/pyproject.toml @@ -0,0 +1,15 @@ +[project] +name = "comfyui-mmaudio" +description = "Nodes to generate audio for vide with: [a/https://github.com/hkchengrex/MMAudio](https://github.com/hkchengrex/MMAudio)" +version = "1.0.3" +license = {file = "LICENSE"} +dependencies = ["accelerate>=0.33.0", "numpy<=1.26.4", "transformers>=4.44.0", "diffusers>=0.25.0", "ftfy>=6.1.1", "opencv-python>=4.7.0.68", "einops>=0.7.0", "bitsandbytes>=0.43.3", "prodigyopt>=1.0", "lion-pytorch>=0.0.6", "safetensors>=0.4.2", "altair>=4.2.2", "toml>=0.10.2", "voluptuous>=0.13.1", "huggingface-hub>=0.24.5", "# for Image utils", "imagesize>=1.4.1", "rich>=13.7.0", "came_pytorch", "matplotlib", "# for T5XXL tokenizer (SD3/FLUX)", "sentencepiece>=0.2.0"] + +[project.urls] +Repository = "https://github.com/kijai/ComfyUI-MMAudio" +# Used by Comfy Registry https://comfyregistry.org + +[tool.comfy] +PublisherId = "kijai" +DisplayName = "ComfyUI-MMAudio" +Icon = "" diff --git a/custom_nodes/ComfyUI-MMAudio/readme.md b/custom_nodes/ComfyUI-MMAudio/readme.md new file mode 100644 index 0000000000000000000000000000000000000000..2ca9f69d4303baa28ca0ceafd71183904450f0f7 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/readme.md @@ -0,0 +1,25 @@ +# ComfyUI nodes to use [MMAudio](https://github.com/hkchengrex/MMAudio) + +## WIP WIP WIP + +https://github.com/user-attachments/assets/9515c0f6-cc5d-4dfe-a642-f841a1a2dba5 + +# Installation +Clone this repo into custom_nodes folder. + +Install dependencies: pip install -r requirements.txt or if you use the portable install, run this in ComfyUI_windows_portable -folder: + +python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-MMAudio\requirements.txt + + +Models are loaded from `ComfyUI/models/mmaudio` + +Safetensors available here: + +https://huggingface.co/Kijai/MMAudio_safetensors/tree/main + +Nvidia bigvganv2 (used with 44k mode) + +https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_512x + +is autodownloaded to `ComfyUI/models/mmaudio/nvidia/bigvgan_v2_44khz_128band_512x` diff --git a/custom_nodes/ComfyUI-MMAudio/requirements.txt b/custom_nodes/ComfyUI-MMAudio/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..7b37832a8565dd8df56c552d8799bed36a913bd2 --- /dev/null +++ b/custom_nodes/ComfyUI-MMAudio/requirements.txt @@ -0,0 +1,8 @@ +librosa>=0.10.1 +torchdiffeq>=0.2.3 +einops>=0.7.0 +timm>=1.0.8 +omegaconf>=2.3.0 +open_clip_torch>=2.29.0 +accelerate +ftfy>=6.1.1 \ No newline at end of file diff --git a/custom_nodes/comfyui-videohelpersuite/.github/workflows/publish.yml b/custom_nodes/comfyui-videohelpersuite/.github/workflows/publish.yml new file mode 100644 index 0000000000000000000000000000000000000000..27bcdb11695ea19286da34cd5f36cf75155a4dbb --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/.github/workflows/publish.yml @@ -0,0 +1,24 @@ +name: Publish to Comfy registry +on: + workflow_dispatch: + push: + branches: + - main + paths: + - "pyproject.toml" + +permissions: + issues: write + +jobs: + publish-node: + name: Publish Custom Node to registry + runs-on: ubuntu-latest + if: ${{ github.repository_owner == 'Kosinkadink' }} + steps: + - name: Check out code + uses: actions/checkout@v4 + - name: Publish Custom Node + uses: Comfy-Org/publish-node-action@v1 + with: + personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }} ## Add your own personal access token to your Github Repository secrets and reference it here. diff --git a/custom_nodes/comfyui-videohelpersuite/.gitignore b/custom_nodes/comfyui-videohelpersuite/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..9b5379e8f93da3909706b0a7d37f5c8ab5e2a2a5 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/.gitignore @@ -0,0 +1,162 @@ +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +share/python-wheels/ +*.egg-info/ +.installed.cfg +*.egg +MANIFEST + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.nox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +*.py,cover +.hypothesis/ +.pytest_cache/ +cover/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py +db.sqlite3 +db.sqlite3-journal + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +.pybuilder/ +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# IPython +profile_default/ +ipython_config.py + +# pyenv +# For a library or package, you might want to ignore these files since the code is +# intended to run in multiple environments; otherwise, check them in: +# .python-version + +# pipenv +# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. +# However, in case of collaboration, if having platform-specific dependencies or dependencies +# having no cross-platform support, pipenv may install dependencies that don't work, or not +# install all needed dependencies. +#Pipfile.lock + +# poetry +# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. +# This is especially recommended for binary packages to ensure reproducibility, and is more +# commonly ignored for libraries. +# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control +#poetry.lock + +# pdm +# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. +#pdm.lock +# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it +# in version control. +# https://pdm.fming.dev/#use-with-ide +.pdm.toml + +# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm +__pypackages__/ + +# Celery stuff +celerybeat-schedule +celerybeat.pid + +# SageMath parsed files +*.sage.py + +# Environments +.env +.venv +env/ +venv/ +ENV/ +env.bak/ +venv.bak/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ +.dmypy.json +dmypy.json + +# Pyre type checker +.pyre/ + +# pytype static type analyzer +.pytype/ + +# Cython debug symbols +cython_debug/ + +# PyCharm +# JetBrains specific template is maintained in a separate JetBrains.gitignore that can +# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore +# and can be added to the global gitignore or merged into this file. For a more nuclear +# option (not recommended) you can uncomment the following to ignore the entire idea folder. +#.idea/ +# VIM swap files +*.swp diff --git a/custom_nodes/comfyui-videohelpersuite/LICENSE b/custom_nodes/comfyui-videohelpersuite/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..f288702d2fa16d3cdf0035b15a9fcbc552cd88e7 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/LICENSE @@ -0,0 +1,674 @@ + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + The GNU General Public License does not permit incorporating your program +into proprietary programs. If your program is a subroutine library, you +may consider it more useful to permit linking proprietary applications with +the library. If this is what you want to do, use the GNU Lesser General +Public License instead of this License. But first, please read +. diff --git a/custom_nodes/comfyui-videohelpersuite/README.md b/custom_nodes/comfyui-videohelpersuite/README.md new file mode 100644 index 0000000000000000000000000000000000000000..753fcecf9979b40b2427869c6ad5bed1b61df482 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/README.md @@ -0,0 +1,110 @@ +# ComfyUI-VideoHelperSuite +Nodes related to video workflows + +## I/O Nodes +### Load Video +Converts a video file into a series of images +- video: The video file to be loaded +- force_rate: Discards or duplicates frames as needed to hit a target frame rate. Disabled by setting to 0. This can be used to quickly match a suggested frame rate like the 8 fps of AnimateDiff. +- force_size: Allows for quick resizing to a number of suggested sizes. Several options allow you to set only width or height and determine the other from aspect ratio. +- frame_load_cap: The maximum number of frames which will be returned. This could also be thought of as the maximum batch size. +- skip_first_frames: How many frames to skip from the start of the video after adjusting for a forced frame rate. By incrementing this number by the frame_load_cap, you can easily process a longer input video in parts. +- select_every_nth: Allows for skipping a number of frames without considering the base frame rate or risking frame duplication. Often useful when working with animated gifs +A path variant of the Load Video node exists that allows loading videos from external paths +![step](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/assets/4284322/b5fc993c-5c9b-4608-afa4-48ae2e1380ef) +![resize](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/assets/4284322/98d2e78e-1c44-443c-a8fe-0dab0b5947f3) +If [Advanced Previews](#advanced-previews) is enabled in the options menu of the web ui, the preview will reflect the current settings on the node. +### Load Image Sequence +Loads all image files from a subfolder. Options are similar to Load Video. +- image_load_cap: The maximum number of images which will be returned. This could also be thought of as the maximum batch size. +- skip_first_images: How many images to skip. By incrementing this number by image_load_cap, you can easily divide a long sequence of images into multiple batches. +- select_every_nth: Allows for skipping a number of images between every returned frame. + +A path variant of Load Image sequence also exists. +### Video Combine +Combines a series of images into an output video +If the optional audio input is provided, it will also be combined into the output video +- frame_rate: How many of the input frames are displayed per second. A higher frame rate means that the output video plays faster and has less duration. This should usually be kept to 8 for AnimateDiff, or matched to the force_rate of a Load Video node. +- loop_count: How many additional times the video should repeat +- filename_prefix: The base file name used for output. + - You can save output to a subfolder: `subfolder/video` + - Like the builtin Save Image node, you can add timestamps. `%date:yyyy-MM-ddThh:mm:ss%` might become 2023-10-31T6:45:25 +- format: The file format to use. Advanced information on configuring or adding additional video formats can be found in the [Video Formats](#video-formats) section. +- pingpong: Causes the input to be played back in the reverse to create a clean loop. +- save_output: Whether the image should be put into the output directory or the temp directory. +Returns: a `VHS_FILENAMES` which consists of a boolean indicating if save_output is enabled and a list of the full filepaths of all generated outputs in the order created. Accordingly `output[1][-1]` will be the most complete output. + +Depending on the format chosen, additional options may become available, including +- crf: Describes the quality of the output video. A lower number gives a higher quality video and a larger file size, while a higher number gives a lower quality video with a smaller size. Scaling varies by codec, but visually lossless output generally occurs around 20. +- save_metadata: Includes a copy of the workflow in the output video which can be loaded by dragging and dropping the video, just like with images. +- pix_fmt: Changes how the pixel data is stored. `yuv420p10le` has higher color quality, but won't work on all devices +### Load Audio +Provides a way to load standalone audio files. +- seek_seconds: An optional start time for the audio file in seconds. + +## Latent/Image Nodes +A number of utility nodes exist for managing latents. For each, there is an equivalent node which works on images. +### Split Batch +Divides the latents into two sets. The first `split_index` latents go to output A and the remainder to output B. If less then `split_index` latents are provided as input, all are passed to output A and output B is empty. +### Merge Batch +Combines two groups of latents into a single output. The order of the output is the latents in A followed by the latents in B. +If the input groups are not the same size, the node provides options for rescaling the latents before merging. +### Select Every Nth +The first of every `select_every_nth` input is passed and the remainder are discarded +### Get Count +### Duplicate Batch + +## Video Previews +Load Video (Upload), Load Video (Path), Load Images (Upload), Load Images (Path) and Video Combine provide animated previews. +Nodes with previews provide additional functionality when right clicked +- Open preview +- Save preview +- Pause preview: Can improve performance with very large videos +- Hide preview: Can improve performance, save space +- Sync preview: Restarts all previews for side-by-side comparisons + +### Advanced Previews +Advanced Previews must be manually enabled by clicking the settings gear next to Queue Prompt and checking the box for VHS Advanced Previews. +If enabled, videos which are displayed in the ui will be converted with ffmpeg on request. This has several benefits +- Previews for Load Video nodes will reflect the settings on the node such as skip_first_frames and frame_load_cap + - This makes it easy to select an exact portion of an input video and sync it with outputs +- It can use substantially less bandwidth if running the server remotely +- It can greatly improve the browser performance by downsizing videos to the in ui resolution, particularly useful with animated gifs +- It allows for previews of videos that would not normally be playable in browser. +- Can be limited to subdirectories of ComyUI if `VHS_STRICT_PATHS` is set as an environment variable. + +This fucntionality is disabled since it comes with several downsides +- There is a delay before videos show in the browser. This delay can become quite large if the input video is long +- The preview videos are lower quality (The original can always be viewed with Right Click -> Open preview) + +## Video Formats +Those familiar with ffmpeg are able to add json files to the video_formats folders to add new output types to Video Combine. +Consider the following example for av1-webm +```json +{ + "main_pass": + [ + "-n", "-c:v", "libsvtav1", + "-pix_fmt", "yuv420p10le", + "-crf", ["crf","INT", {"default": 23, "min": 0, "max": 100, "step": 1}] + ], + "audio_pass": ["-c:a", "libopus"], + "extension": "webm", + "environment": {"SVT_LOG": "1"} +} +``` +Most configuration takes place in `main_pass`, which is a list of arguments that are passed to ffmpeg. +- `"-n"` designates that the command should fail if a file of the same name already exists. This should never happen, but if some bug were to occur, it would ensure other files aren't overwritten. +- `"-c:v", "libsvtav1"` designates that the video should be encoded with an av1 codec using the new SVT-AV1 encoder. SVT-AV1 is much faster than libaom-av1, but may not exist in older versions of ffmpeg. Alternatively, av1_nvenc could be used for gpu encoding with newer nvidia cards. +- `"-pix_fmt", "yuv420p10le"` designates the standard pixel format with 10-bit color. It's important that some pixel format be specified to ensure a nonconfigurable input pix_fmt isn't used. + +`audio pass` contains a list of arguments which are passed to ffmpeg when audio is passed into Video Combine + +`extension` designates both the file extension and the container format that is used. If some of the above options are omitted from `main_pass` it can affect what default options are chosen. +`environment` can optionally be provided to set environment variables during execution. For av1 it's used to reduce the verbosity of logging so that only major errors are displayed. +`input_color_depth` effects the format in which pixels are passed to the ffmpeg subprocess. Current valid options are `8bit` and `16bit`. The later will produce higher quality output, but is experimental. + +Fields can be exposed in the webui as a widget using a format similar to what is used in the creation of custom nodes. In the above example, the argument for `-crf` will be exposed as a format widget in the webui. Format widgets are a list of up to 3 terms +- The name of the widget that will be displayed in the web ui +- Either a primitive such as "INT" or "BOOLEAN", or a list of string options +- A dictionary of options diff --git a/custom_nodes/comfyui-videohelpersuite/__init__.py b/custom_nodes/comfyui-videohelpersuite/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9fa995f0ac7e296a62bcba9df68b5250a0f7d041 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/__init__.py @@ -0,0 +1,9 @@ +from .videohelpersuite.nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS +import folder_paths +from .videohelpersuite.server import server +from .videohelpersuite import documentation +from .videohelpersuite import latent_preview + +WEB_DIRECTORY = "./web" +__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS", "WEB_DIRECTORY"] +documentation.format_descriptions(NODE_CLASS_MAPPINGS) diff --git a/custom_nodes/comfyui-videohelpersuite/pyproject.toml b/custom_nodes/comfyui-videohelpersuite/pyproject.toml new file mode 100644 index 0000000000000000000000000000000000000000..39793ee53bd813c5188a2f65836a72bf541b4d9c --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/pyproject.toml @@ -0,0 +1,15 @@ +[project] +name = "comfyui-videohelpersuite" +description = "Nodes related to video workflows" +version = "1.7.9" +license = { file = "LICENSE" } +dependencies = ["opencv-python", "imageio-ffmpeg"] + +[project.urls] +Repository = "https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite" + +# Used by Comfy Registry https://comfyregistry.org +[tool.comfy] +PublisherId = "kosinkadink" +DisplayName = "ComfyUI-VideoHelperSuite" +Icon = "" diff --git a/custom_nodes/comfyui-videohelpersuite/requirements.txt b/custom_nodes/comfyui-videohelpersuite/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..4fa34aa21b85c4b974e2a2b6891eae5fd6dd4164 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/requirements.txt @@ -0,0 +1,2 @@ +opencv-python +imageio-ffmpeg diff --git a/custom_nodes/comfyui-videohelpersuite/testframework/README.md b/custom_nodes/comfyui-videohelpersuite/testframework/README.md new file mode 100644 index 0000000000000000000000000000000000000000..93cd503563080c9c7d785d419bef7dddc4c11cd8 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/testframework/README.md @@ -0,0 +1,5 @@ +Code to automate execution of the tests and evaluate the results. +Distributed as a `custom node`, and can be installed by copying or simlinking to the `custom_nodes` directory. +Requires that ffprobe be available and added to the path. Note that imageio-ffmpeg does not bundle ffprobe. + +When installed, it adds a new sidebar tab to automate running one, or a folder of tests. This requires that the `Use new menu and workflow management` setting not be disabled diff --git a/custom_nodes/comfyui-videohelpersuite/testframework/__init__.py b/custom_nodes/comfyui-videohelpersuite/testframework/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c75b11c164596ef873898dd843700c7164cc0932 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/testframework/__init__.py @@ -0,0 +1,6 @@ +from . import server +NODE_CLASS_MAPPINGS = {} +NODE_DISPLAY_NAME_MAPPINGS = {} + +WEB_DIRECTORY = "./web" +__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS", "WEB_DIRECTORY"] diff --git a/custom_nodes/comfyui-videohelpersuite/testframework/server.py b/custom_nodes/comfyui-videohelpersuite/testframework/server.py new file mode 100644 index 0000000000000000000000000000000000000000..a9c20ea3eab31bab288d6089a89410ff409d43ad --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/testframework/server.py @@ -0,0 +1,60 @@ +import subprocess +import json +import os +import torch +import shutil + +import server +import folder_paths + +web = server.web + +@server.PromptServer.instance.routes.post("/VHS_test") +async def test(request): + try: + req_data = await request.json() + output = req_data['output']['gifs'][0] + filename = output['filename'] + typ = output['type'] + base_args = ["ffprobe", "-v", "error", '-count_packets', "-show_entries", "stream", "-of", "json"] + video = folder_paths.get_annotated_filepath(f'{filename} [{typ}]') + vprobe = json.loads(subprocess.run(base_args + ['-select_streams', 'v:0', video], + capture_output=True, check=True).stdout)['streams'][0] + aprobe = json.loads(subprocess.run(base_args + ['-select_streams', 'a:0', video], + capture_output=True, check=True).stdout)['streams'] + probe = {'video': vprobe} + if len(aprobe) > 0: + probe['audio'] = aprobe[0] + errors = [] + compare = None + for test in req_data['tests']: + if test['type'] == 'compare': + compare = test + continue + key = test['key'] + expected = test['value'] + actual = probe[test['type']][key] + if expected != actual: + #Consider always dumping type? + errors.append(f'{key}: {expected} != {actual}') + if len(errors) == 0 and compare is not None: + if not os.path.exists(compare['filename']): + os.makedirs(os.path.split(compare['filename'])[0], exist_ok=True) + shutil.copy(video, compare['filename']) + print("Missing comparison file has been initialized from output:", os.path.abspath(compare['filename'])) + else: + #NOTE: This does not include the full memory optimizations of VHS + #Tests should be small + #TODO: Figure out way to do opacity comparison. May need to do blending in python + #(easy, but slower and more memory intensive) + diff = subprocess.run(['ffmpeg', '-v', 'error', '-i', video, '-i', compare['filename'], '-filter_complex', 'blend=all_mode=grainextract', '-pix_fmt', 'rgb24', '-f', 'rawvideo', '-'], stdout=subprocess.PIPE, check=True).stdout + diff = torch.frombuffer(diff, dtype=torch.uint8).to(dtype=torch.float32).div_(255) + #diff = diff.reshape((-1,4)) + d = (diff-0.5).abs().sum()/diff.size(0) + if d > compare['tolerance']: + errors.append(f'Similarity is outside specified tolerance: {d}') + else: + print('d:', d) + return web.json_response(errors) + except Exception as e: + return web.json_response(str(e)) diff --git a/custom_nodes/comfyui-videohelpersuite/testframework/web/js/testRunner.js b/custom_nodes/comfyui-videohelpersuite/testframework/web/js/testRunner.js new file mode 100644 index 0000000000000000000000000000000000000000..521a2204b9a2a3202441818461231872ea8a15c8 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/testframework/web/js/testRunner.js @@ -0,0 +1,96 @@ + +import {app} from "../../../scripts/app.js"; +import {api} from "../../../scripts/api.js"; + +let watched_nodes = {} +let resolve = undefined +let testURL = api.apiURL("/VHS_test") +let errors = [] +api.addEventListener("executed", async function ({detail}) { + if (watched_nodes && watched_nodes[detail?.node]) { + if (detail?.output?.unfinished_batch) { + return + } + let requestBody = {tests: watched_nodes[detail.node], output: detail.output} + try { + let req = await fetch(api.apiURL("/VHS_test"), + {method: "POST", body: JSON.stringify(requestBody)}); + let testResult = await req.json() + if (testResult.length != 0) { + errors.push(testResult) + } + } catch(e) { + errors.push(e) + } + if (!(watched_nodes.length -= 1)) { + resolve() + } + } +}); + +const workflowService = app.extensionManager.workflow + +async function runTest(file) { + if (!file?.name?.endsWith(".json")) { + return false + } + let workflow = JSON.parse(await file.text()) + await app.loadGraphData(workflow) + //NOTE: API is not used so workflow data is actually processed + watched_nodes = workflow.tests + errors = [] + let p = new Promise((r) => resolve = r) + await app.queuePrompt() + //block until execution completes + await p + watched_nodes = {} + if (errors.length > 0) { + app.ui.dialog.show("Failed " + errors.length + " tests:\n" + errors) + return true + } + await workflowService.closeWorkflow(workflowService.activeWorkflow, {warnIfUnsaved: false}) + return false +} +let iconOverride = document.createElement("style") +iconOverride.innerHTML = `.VHSTestIcon:before {content: '🧪';}` +document.body.append(iconOverride) + +let testSidebar = {id: 'VHStest', title: 'VHS Test', icon: 'VHSTestIcon', type: 'custom', + render: (e) => { + e.innerHTML = `Select a folder containing tests + + Or select a single test + + ` + + const folderInput = e.children[0] + const fileInput = e.children[1] + Object.assign(folderInput, { + type: "file", + webkitdirectory: true, + onchange: async function() { + const startTime = Date.now() + let failedTests = false + for(const file of this.files) { + failedTests ||= await runTest(file) + } + this.value="" + if (!failedTests) { + console.log("All tests passed in " + ((Date.now() - startTime)/1000) + "s") + } + }, + }); + Object.assign(fileInput, { + type: "file", + accept: ".json", + onchange: async function() { + if (this.files.length) { + if(!(await runTest(this.files[0]))) { + console.log("Test complete") + } + this.value="" + } + }, + }); + }} +app.extensionManager.registerSidebarTab(testSidebar) diff --git a/custom_nodes/comfyui-videohelpersuite/tests/README.md b/custom_nodes/comfyui-videohelpersuite/tests/README.md new file mode 100644 index 0000000000000000000000000000000000000000..ec025a7d169c7783f34523c63a0a6ae9a2e00fdb --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/README.md @@ -0,0 +1 @@ +Workflows for automated testing of VHS. Most include an additional tests key to check the properties or perform comparisons on node outputs diff --git a/custom_nodes/comfyui-videohelpersuite/tests/audio.json b/custom_nodes/comfyui-videohelpersuite/tests/audio.json new file mode 100644 index 0000000000000000000000000000000000000000..cae3d766731f7fe3ae4866f6fa8c0cc86a2ca46d --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/audio.json @@ -0,0 +1,313 @@ +{ + "id": "07b812b5-5037-4878-90bc-32d3a1f36619", + "revision": 0, + "last_node_id": 7, + "last_link_id": 5, + "nodes": [ + { + "id": 5, + "type": "VHS_VideoCombine", + "pos": [ + 732, + -23 + ], + "size": [ + 210, + 334 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 2 + }, + { + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 3 + }, + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/webm", + "pix_fmt": "yuv420p", + "crf": 20, + "save_metadata": true, + "trim_to_audio": false, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": {} + } + } + }, + { + "id": 6, + "type": "VHS_VideoCombine", + "pos": [ + 503, + 363 + ], + "size": [ + 210, + 334 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 4 + }, + { + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 5 + }, + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/h264-mp4", + "pix_fmt": "yuv420p", + "crf": 19, + "save_metadata": true, + "trim_to_audio": false, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": {} + } + } + }, + { + "id": 4, + "type": "VHS_LoadVideoPath", + "pos": [ + 29, + 16 + ], + "size": [ + 221.27618408203125, + 413.1552734375 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 2, + 4 + ] + }, + { + "name": "frame_count", + "type": "INT", + "links": null + }, + { + "name": "audio", + "type": "AUDIO", + "links": [ + 3 + ] + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideoPath" + }, + "widgets_values": { + "video": "input/bigbuckbunny.mp4", + "force_rate": 8, + "custom_width": 0, + "custom_height": 0, + "frame_load_cap": 30, + "skip_first_frames": 0, + "select_every_nth": 1, + "format": "AnimateDiff", + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "input/bigbuckbunny.mp4", + "type": "path", + "format": "video/mp4", + "force_rate": 8, + "custom_width": 0, + "custom_height": 0, + "frame_load_cap": 30, + "skip_first_frames": 0, + "select_every_nth": 1 + } + } + } + }, + { + "id": 7, + "type": "VHS_LoadAudio", + "pos": [ + 83, + 564 + ], + "size": [ + 218.93820190429688, + 126 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [], + "outputs": [ + { + "name": "audio", + "type": "AUDIO", + "links": [ + 5 + ] + }, + { + "name": "duration", + "type": "FLOAT", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_LoadAudio" + }, + "widgets_values": { + "audio_file": "input/bigbuckbunny.mp4", + "seek_seconds": 0, + "duration": 5 + } + } + ], + "links": [ + [ + 2, + 4, + 0, + 5, + 0, + "IMAGE" + ], + [ + 3, + 4, + 2, + 5, + 1, + "AUDIO" + ], + [ + 4, + 4, + 0, + 6, + 0, + "IMAGE" + ], + [ + 5, + 7, + 0, + 6, + 1, + "AUDIO" + ] + ], + "groups": [], + "config": {}, + "extra": { + "frontendVersion": "1.25.0", + "VHS_latentpreview": true, + "VHS_latentpreviewrate": 0, + "VHS_MetadataImage": true, + "VHS_KeepIntermediate": true + }, + "version": 0.4, + "tests": { + "5": [{"type": "audio", "key": "nb_read_packets", "value": "177"}], + "6": [{"type": "audio", "key": "nb_read_packets", "value": "176"}], + "length": 2 + } +} \ No newline at end of file diff --git a/custom_nodes/comfyui-videohelpersuite/tests/batch4x4.json b/custom_nodes/comfyui-videohelpersuite/tests/batch4x4.json new file mode 100644 index 0000000000000000000000000000000000000000..34b9647329642a9e5ca233a60b8bbf2ca8f72aca --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/batch4x4.json @@ -0,0 +1,535 @@ +{ + "last_node_id": 18, + "last_link_id": 28, + "nodes": [ + { + "id": 16, + "type": "SolidMask", + "pos": [ + 38, + 1066 + ], + "size": { + "0": 315, + "1": 106 + }, + "flags": {}, + "order": 0, + "mode": 0, + "outputs": [ + { + "name": "MASK", + "type": "MASK", + "links": [ + 18 + ], + "slot_index": 0, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "SolidMask" + }, + "widgets_values": [ + 1, + 512, + 512 + ] + }, + { + "id": 15, + "type": "ImageCompositeMasked", + "pos": [ + 412.0800030517579, + 591.4099975585939 + ], + "size": { + "0": 315, + "1": 146 + }, + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "name": "destination", + "type": "IMAGE", + "link": 16 + }, + { + "name": "source", + "type": "IMAGE", + "link": 17 + }, + { + "name": "mask", + "type": "MASK", + "link": 18 + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 19 + ], + "slot_index": 0, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "ImageCompositeMasked" + }, + "widgets_values": [ + 0, + 0, + false + ] + }, + { + "id": 18, + "type": "VHS_BatchManager", + "pos": [ + 329.8000881958008, + -87.40008056640622 + ], + "size": { + "0": 315, + "1": 58 + }, + "flags": {}, + "order": 1, + "mode": 0, + "outputs": [ + { + "name": "VHS_BatchManager", + "type": "VHS_BatchManager", + "links": [ + 24, + 25, + 26, + 27 + ], + "slot_index": 0, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_BatchManager" + }, + "widgets_values": { + "frames_per_batch": 4, + "count": 2 + } + }, + { + "id": 10, + "type": "VHS_LoadVideoPath", + "pos": [ + 24, + 102 + ], + "size": [ + 320, + 420.7188019966722 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": 25, + "slot_index": 0 + }, + { + "name": "vae", + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 16, + 22 + ], + "slot_index": 0, + "shape": 3 + }, + { + "name": "frame_count", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "audio", + "type": "AUDIO", + "links": [ + 28 + ], + "slot_index": 2, + "shape": 3 + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideoPath" + }, + "widgets_values": { + "video": "input/bigbuckbunny.mp4", + "force_rate": 8, + "force_size": "512x?", + "custom_width": 512, + "custom_height": 512, + "frame_load_cap": 16, + "skip_first_frames": 64, + "select_every_nth": 1, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 16, + "skip_first_frames": 64, + "force_rate": 8, + "filename": "input/bigbuckbunny.mp4", + "type": "path", + "format": "video/mp4", + "select_every_nth": 1, + "force_size": "512x?" + } + } + } + }, + { + "id": 14, + "type": "VHS_LoadVideoPath", + "pos": [ + 26, + 598 + ], + "size": [ + 320, + 444.7188019966722 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": 26, + "slot_index": 0 + }, + { + "name": "vae", + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 17 + ], + "slot_index": 0, + "shape": 3 + }, + { + "name": "frame_count", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "audio", + "type": "VHS_AUDIO", + "links": [], + "slot_index": 2, + "shape": 3 + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideoPath" + }, + "widgets_values": { + "video": "input/bigbuckbunny.mp4", + "force_rate": 8, + "force_size": "Custom Width", + "custom_width": 384, + "custom_height": 512, + "frame_load_cap": 16, + "skip_first_frames": 64, + "select_every_nth": 1, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 16, + "skip_first_frames": 64, + "force_rate": 8, + "filename": "input/bigbuckbunny.mp4", + "type": "path", + "format": "video/mp4", + "select_every_nth": 1, + "force_size": "512x?" + } + } + } + }, + { + "id": 11, + "type": "VHS_VideoCombine", + "pos": [ + 762, + 646 + ], + "size": [ + 320, + 492.75 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 19 + }, + { + "name": "audio", + "type": "AUDIO", + "link": 28 + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": 27, + "slot_index": 2 + }, + { + "name": "vae", + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/h264-mp4", + "pix_fmt": "yuv420p", + "crf": 19, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00005-audio.mp4", + "subfolder": "", + "type": "temp", + "format": "video/h264-mp4", + "frame_rate": 8 + } + } + } + }, + { + "id": 17, + "type": "VHS_VideoCombine", + "pos": [ + 756, + 100 + ], + "size": [ + 320, + 492.75 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 22 + }, + { + "name": "audio", + "type": "VHS_AUDIO", + "link": null + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": 24, + "slot_index": 2 + }, + { + "name": "vae", + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/webm", + "pix_fmt": "yuv420p", + "crf": 20, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00004.webm", + "subfolder": "", + "type": "temp", + "format": "video/webm", + "frame_rate": 8 + } + } + } + } + ], + "links": [ + [ + 16, + 10, + 0, + 15, + 0, + "IMAGE" + ], + [ + 17, + 14, + 0, + 15, + 1, + "IMAGE" + ], + [ + 18, + 16, + 0, + 15, + 2, + "MASK" + ], + [ + 19, + 15, + 0, + 11, + 0, + "IMAGE" + ], + [ + 22, + 10, + 0, + 17, + 0, + "IMAGE" + ], + [ + 24, + 18, + 0, + 17, + 2, + "VHS_BatchManager" + ], + [ + 25, + 18, + 0, + 10, + 0, + "VHS_BatchManager" + ], + [ + 26, + 18, + 0, + 14, + 0, + "VHS_BatchManager" + ], + [ + 27, + 18, + 0, + 11, + 2, + "VHS_BatchManager" + ], + [ + 28, + 10, + 2, + 11, + 1, + "VHS_AUDIO" + ] + ], + "groups": [], + "config": {}, + "extra": {}, + "version": 0.4, + "tests": { + "17": [{"type": "video", "key": "nb_read_packets", "value": "16"}], + "11": [{"type": "video", "key": "nb_read_packets", "value": "16"}, + {"type": "compare", "filename": "custom_nodes/ComfyUI-VideoHelperSuite/tests/outputs/batch.mp4", "tolerance": 0.02} + ], + "length": 1 + } +} diff --git a/custom_nodes/comfyui-videohelpersuite/tests/converted-format-input.json b/custom_nodes/comfyui-videohelpersuite/tests/converted-format-input.json new file mode 100644 index 0000000000000000000000000000000000000000..e9722580df540b89a4b9185260a8de180ddd63aa --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/converted-format-input.json @@ -0,0 +1,253 @@ +{ + "last_node_id": 19, + "last_link_id": 20, + "nodes": [ + { + "id": 18, + "type": "PrimitiveNode", + "pos": [ + 318, + 618 + ], + "size": [ + 210, + 82 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [], + "outputs": [ + { + "name": "INT", + "type": "INT", + "links": [ + 18 + ], + "widget": { + "name": "crf" + } + } + ], + "properties": { + "Run widget replace on values": false + }, + "widgets_values": [ + 60, + "fixed" + ] + }, + { + "id": 11, + "type": "LoadImage", + "pos": [ + 260.4530029296875, + 233.2003173828125 + ], + "size": [ + 315, + 314 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 19 + ], + "slot_index": 0 + }, + { + "name": "MASK", + "type": "MASK", + "links": null + } + ], + "properties": { + "Node name for S&R": "LoadImage" + }, + "widgets_values": [ + "example.png", + "image" + ] + }, + { + "id": 17, + "type": "VHS_VideoCombine", + "pos": [ + 733.3749389648438, + 338.28924560546875 + ], + "size": [ + 222.91415405273438, + 522.9141845703125 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 19 + }, + { + "name": "audio", + "type": "AUDIO", + "link": null, + "shape": 7 + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null, + "shape": 7 + }, + { + "name": "vae", + "type": "VAE", + "link": null, + "shape": 7 + }, + { + "name": "crf", + "type": "INT", + "link": 18, + "widget": { + "name": "crf" + } + }, + { + "name": "pix_fmt", + "type": [ + "yuv420p", + "yuv420p10le" + ], + "link": 20, + "widget": { + "name": "pix_fmt" + } + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/h264-mp4", + "pix_fmt": "yuv420p10le", + "crf": 60, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00001.mp4", + "subfolder": "", + "type": "temp", + "format": "video/h264-mp4", + "frame_rate": 8 + }, + "muted": false + } + } + }, + { + "id": 19, + "type": "PrimitiveNode", + "pos": [ + 300, + 760 + ], + "size": [ + 290, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [], + "outputs": [ + { + "name": "COMBO", + "type": "COMBO", + "links": [ + 20 + ], + "widget": { + "name": "pix_fmt" + } + } + ], + "properties": { + "Run widget replace on values": false + }, + "widgets_values": [ + "yuv420p10le", + "fixed", + "" + ] + } + ], + "links": [ + [ + 18, + 18, + 0, + 17, + 4, + "INT" + ], + [ + 19, + 11, + 0, + 17, + 0, + "IMAGE" + ], + [ + 20, + 19, + 0, + 17, + 5, + [ + "yuv420p", + "yuv420p10le" + ] + ] + ], + "groups": [], + "config": {}, + "extra": { + "ds": { + "scale": 0.8264462809917354, + "offset": [ + 45.8650452880864, + -157.46987175292935 + ] + } + }, + "version": 0.4, + "tests": { + "17": [{"type": "video", "key": "pix_fmt", "value": "yuv420p10le"} + ], + "length": 1 + } +} diff --git a/custom_nodes/comfyui-videohelpersuite/tests/converted-input.json b/custom_nodes/comfyui-videohelpersuite/tests/converted-input.json new file mode 100644 index 0000000000000000000000000000000000000000..41a93a049ea95ea9ddda92ed3a5152b55581938f --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/converted-input.json @@ -0,0 +1,307 @@ +{ + "last_node_id": 20, + "last_link_id": 23, + "nodes": [ + { + "id": 18, + "type": "VHS_VideoInfoLoaded", + "pos": [ + 424, + 427 + ], + "size": { + "0": 304.79998779296875, + "1": 106 + }, + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "link": 19 + } + ], + "outputs": [ + { + "name": "fps🟦", + "type": "FLOAT", + "links": [ + 20 + ], + "slot_index": 0, + "shape": 3 + }, + { + "name": "frame_count🟦", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "duration🟦", + "type": "FLOAT", + "links": null, + "shape": 3 + }, + { + "name": "width🟦", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "height🟦", + "type": "INT", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_VideoInfoLoaded" + }, + "widgets_values": {} + }, + { + "id": 17, + "type": "VHS_VideoCombine", + "pos": [ + 783, + 223 + ], + "size": [ + 315, + 286 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 21 + }, + { + "name": "audio", + "type": "VHS_AUDIO", + "link": null + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "frame_rate", + "type": "FLOAT", + "link": 20, + "widget": { + "name": "frame_rate" + } + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/h264-mp4", + "pix_fmt": "yuv420p", + "crf": 19, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00001.mp4", + "subfolder": "", + "type": "temp", + "format": "video/h264-mp4" + } + } + } + }, + { + "id": 16, + "type": "VHS_LoadVideoPath", + "pos": [ + 96, + 230 + ], + "size": [ + 315, + 449.5 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "frame_load_cap", + "type": "INT", + "link": 23, + "slot_index": 1, + "widget": { + "name": "frame_load_cap" + } + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 21 + ], + "slot_index": 0, + "shape": 3 + }, + { + "name": "frame_count", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "audio", + "type": "VHS_AUDIO", + "links": null, + "shape": 3 + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": [ + 19 + ], + "slot_index": 3, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideoPath" + }, + "widgets_values": { + "video": "input/leader.webm", + "force_rate": 0, + "force_size": "Disabled", + "custom_width": 512, + "custom_height": 512, + "frame_load_cap": 64, + "skip_first_frames": 0, + "select_every_nth": 1, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 64, + "skip_first_frames": 0, + "force_rate": 0, + "filename": "input/leader.webm", + "type": "path", + "format": "video/webm", + "select_every_nth": 1 + } + } + } + }, + { + "id": 20, + "type": "PrimitiveNode", + "pos": [ + -240, + 170 + ], + "size": { + "0": 210, + "1": 80 + }, + "flags": {}, + "order": 0, + "mode": 0, + "outputs": [ + { + "name": "INT", + "type": "INT", + "links": [ + 23 + ], + "widget": { + "name": "frame_load_cap" + } + } + ], + "properties": { + "Run widget replace on values": false + }, + "widgets_values": [ + 64, + "fixed" + ] + } + ], + "links": [ + [ + 19, + 16, + 3, + 18, + 0, + "VHS_VIDEOINFO" + ], + [ + 20, + 18, + 0, + 17, + 3, + "FLOAT" + ], + [ + 21, + 16, + 0, + 17, + 0, + "IMAGE" + ], + [ + 23, + 20, + 0, + 16, + 1, + "INT" + ] + ], + "groups": [], + "config": {}, + "extra": {}, + "version": 0.4, + "tests": { + "17": [{"type": "video", "key": "width", "value": 1440}, + {"type": "video", "key": "height", "value": 1080}, + {"type": "video", "key": "nb_read_packets", "value": "64"} + ], + "length": 1 + } +} diff --git a/custom_nodes/comfyui-videohelpersuite/tests/loop.json b/custom_nodes/comfyui-videohelpersuite/tests/loop.json new file mode 100644 index 0000000000000000000000000000000000000000..aa5dd4ecdaac29d04b71dbf3bbc8349f2d1210af --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/loop.json @@ -0,0 +1,178 @@ +{ + "last_node_id": 3, + "last_link_id": 1, + "nodes": [ + { + "id": 1, + "type": "VHS_LoadVideo", + "pos": { + "0": 54, + "1": 89 + }, + "size": [ + 260, + 460 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 1 + ], + "slot_index": 0, + "shape": 3 + }, + { + "name": "frame_count", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "audio", + "type": "VHS_AUDIO", + "links": null, + "shape": 3 + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideo" + }, + "widgets_values": { + "video": "leader.webm", + "force_rate": 8, + "force_size": "Disabled", + "custom_width": 304, + "custom_height": 312, + "frame_load_cap": 16, + "skip_first_frames": 1, + "select_every_nth": 1, + "choose video to upload": "image", + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 16, + "skip_first_frames": 1, + "force_rate": 8, + "filename": "leader.webm", + "type": "input", + "format": "video/mp4", + "force_size": "410.4x?", + "select_every_nth": 1 + } + } + } + }, + { + "id": 3, + "type": "VHS_VideoCombine", + "pos": { + "0": 629, + "1": 222 + }, + "size": [ + 320, + 550 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 1 + }, + { + "name": "audio", + "type": "AUDIO", + "link": null + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 1, + "filename_prefix": "AnimateDiff", + "format": "video/h264-mp4", + "pix_fmt": "yuv420p", + "crf": 19, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00005.mp4", + "subfolder": "", + "type": "temp", + "format": "video/h264-mp4", + "frame_rate": 8 + } + } + } + } + ], + "links": [ + [ + 1, + 1, + 0, + 3, + 0, + "IMAGE" + ] + ], + "groups": [], + "config": {}, + "extra": {}, + "version": 0.4, + "tests": { + "3": [{"type": "video", "key": "nb_read_packets", "value": "32"}], + "length": 1 + } +} diff --git a/custom_nodes/comfyui-videohelpersuite/tests/old-prores.json b/custom_nodes/comfyui-videohelpersuite/tests/old-prores.json new file mode 100644 index 0000000000000000000000000000000000000000..17e77c1547663d39b5356ac15d74f807c69dc671 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/old-prores.json @@ -0,0 +1,423 @@ +{ + "id": "bc7448ba-118f-4226-95ab-32227992f954", + "revision": 0, + "last_node_id": 6, + "last_link_id": 4, + "nodes": [ + { + "id": 1, + "type": "VHS_LoadVideo", + "pos": [ + 54, + 89 + ], + "size": [ + 245.1999969482422, + 492.7751770019531 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 1, + 2, + 3, + 4 + ] + }, + { + "name": "frame_count", + "type": "INT", + "links": null + }, + { + "name": "audio", + "type": "AUDIO", + "links": null + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideo" + }, + "widgets_values": { + "video": "leader.webm", + "force_rate": 8, + "custom_width": 512, + "custom_height": 0, + "frame_load_cap": 64, + "skip_first_frames": 1, + "select_every_nth": 1, + "format": "AnimateDiff", + "choose video to upload": "image", + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 64, + "skip_first_frames": 0, + "force_rate": 8, + "filename": "leader.webm", + "type": "input", + "format": "video/mp4", + "select_every_nth": 1 + } + } + } + }, + { + "id": 4, + "type": "VHS_VideoCombine", + "pos": [ + 630.9500122070312, + 136.90997314453125 + ], + "size": [ + 315, + 497.25 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 2 + }, + { + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": null + }, + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/ProRes", + "profile": "2", + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00004.mov", + "subfolder": "", + "type": "temp", + "format": "video/ProRes", + "frame_rate": 8, + "workflow": "AnimateDiff_00004.png" + } + } + } + }, + { + "id": 3, + "type": "VHS_VideoCombine", + "pos": [ + 635.0499267578125, + -407.1000061035156 + ], + "size": [ + 315, + 497.25 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 1 + }, + { + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": null + }, + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/ProRes", + "profile": "1", + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00001.mov", + "subfolder": "", + "type": "temp", + "format": "video/ProRes", + "frame_rate": 8, + "workflow": "AnimateDiff_00001.png" + } + } + } + }, + { + "id": 5, + "type": "VHS_VideoCombine", + "pos": [ + 974.6401977539062, + -409.33984375 + ], + "size": [ + 315, + 497.25 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 3 + }, + { + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": null + }, + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/ProRes", + "profile": "3", + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00002.mov", + "subfolder": "", + "type": "temp", + "format": "video/ProRes", + "frame_rate": 8, + "workflow": "AnimateDiff_00002.png" + } + } + } + }, + { + "id": 6, + "type": "VHS_VideoCombine", + "pos": [ + 968.7000122070312, + 138.7698974609375 + ], + "size": [ + 315, + 497.25 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 4 + }, + { + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": null + }, + { + "name": "meta_batch", + "shape": 7, + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "shape": 7, + "type": "VAE", + "link": null + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/ProRes", + "profile": "4", + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00003.mov", + "subfolder": "", + "type": "temp", + "format": "video/ProRes", + "frame_rate": 8 + } + } + } + } + ], + "links": [ + [ + 1, + 1, + 0, + 3, + 0, + "IMAGE" + ], + [ + 2, + 1, + 0, + 4, + 0, + "IMAGE" + ], + [ + 3, + 1, + 0, + 5, + 0, + "IMAGE" + ], + [ + 4, + 1, + 0, + 6, + 0, + "IMAGE" + ] + ], + "groups": [], + "config": {}, + "extra": { + "frontendVersion": "1.17.0", + "VHS_latentpreview": true, + "VHS_latentpreviewrate": 0, + "VHS_MetadataImage": true, + "VHS_KeepIntermediate": true + }, + "version": 0.4, + "tests": { + "6": [{"type": "video", "key": "pix_fmt", "value": "yuv444p12le"}], + "length": 1 + } +} \ No newline at end of file diff --git a/custom_nodes/comfyui-videohelpersuite/tests/old-vae-conversion.json b/custom_nodes/comfyui-videohelpersuite/tests/old-vae-conversion.json new file mode 100644 index 0000000000000000000000000000000000000000..e751918b6473cae2f38536d8433bc40b2283a3f6 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/old-vae-conversion.json @@ -0,0 +1,229 @@ +{ + "last_node_id": 10, + "last_link_id": 9, + "nodes": [ + { + "id": 7, + "type": "VHS_VideoCombine", + "pos": [ + 746, + 309 + ], + "size": [ + 320, + 468.7188019966722 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "LATENT", + "link": 9 + }, + { + "name": "audio", + "type": "AUDIO", + "link": null + }, + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "type": "VAE", + "link": 8 + } + ], + "outputs": [ + { + "name": "Filenames", + "type": "VHS_FILENAMES", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/webm", + "crf": 20, + "save_metadata": true, + "pingpong": false, + "save_output": false, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "filename": "AnimateDiff_00001.webm", + "subfolder": "", + "type": "temp", + "format": "video/webm", + "frame_rate": 8 + } + } + } + }, + { + "id": 6, + "type": "VHS_LoadVideoPath", + "pos": [ + 171, + 239 + ], + "size": [ + 320, + 420.7188019966722 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "name": "meta_batch", + "type": "VHS_BatchManager", + "link": null + }, + { + "name": "vae", + "type": "VAE", + "link": 7, + "slot_index": 1 + } + ], + "outputs": [ + { + "name": "LATENT", + "type": "LATENT", + "links": [ + 9 + ], + "slot_index": 0, + "shape": 3 + }, + { + "name": "frame_count", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "audio", + "type": "AUDIO", + "links": [], + "slot_index": 2, + "shape": 3 + }, + { + "name": "video_info", + "type": "VHS_VIDEOINFO", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideoPath" + }, + "widgets_values": { + "video": "input/bigbuckbunny.mp4", + "force_rate": 8, + "force_size": "Disabled", + "custom_width": 512, + "custom_height": 512, + "frame_load_cap": 64, + "skip_first_frames": 0, + "select_every_nth": 1, + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 64, + "skip_first_frames": 0, + "force_rate": 8, + "filename": "input/bigbuckbunny.mp4", + "type": "path", + "format": "video/mp4", + "select_every_nth": 1 + } + } + } + }, + { + "id": 10, + "type": "VAELoader", + "pos": [ + -42, + 88 + ], + "size": { + "0": 315, + "1": 58 + }, + "flags": {}, + "order": 0, + "mode": 0, + "outputs": [ + { + "name": "VAE", + "type": "VAE", + "links": [ + 7, + 8 + ], + "shape": 3, + "slot_index": 0 + } + ], + "properties": { + "Node name for S&R": "VAELoader" + }, + "widgets_values": [ + "taesd" + ] + } + ], + "links": [ + [ + 7, + 10, + 0, + 6, + 1, + "VAE" + ], + [ + 8, + 10, + 0, + 7, + 3, + "VAE" + ], + [ + 9, + 6, + 0, + 7, + 0, + "LATENT" + ] + ], + "groups": [], + "config": {}, + "extra": {}, + "version": 0.4, + "tests": { + "7": [{"type": "video", "key": "width", "value": 1920}, + {"type": "video", "key": "height", "value": 1080}, + {"type": "video", "key": "nb_read_packets", "value": "64"} + ], + "length": 1 + } +} diff --git a/custom_nodes/comfyui-videohelpersuite/tests/simple.json b/custom_nodes/comfyui-videohelpersuite/tests/simple.json new file mode 100644 index 0000000000000000000000000000000000000000..4d7f8ddd68a24cc90519ecd127daae9d8ab81b85 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/tests/simple.json @@ -0,0 +1,134 @@ +{ + "last_node_id": 3, + "last_link_id": 1, + "nodes": [ + { + "id": 1, + "type": "VHS_LoadVideo", + "pos": [ + 54, + 89 + ], + "size": [ + 235.1999969482422, + 384.56999829610186 + ], + "flags": {}, + "order": 0, + "mode": 0, + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 1 + ], + "shape": 3, + "slot_index": 0 + }, + { + "name": "frame_count", + "type": "INT", + "links": null, + "shape": 3 + }, + { + "name": "audio", + "type": "VHS_AUDIO", + "links": null, + "shape": 3 + } + ], + "properties": { + "Node name for S&R": "VHS_LoadVideo" + }, + "widgets_values": { + "video": "leader.webm", + "force_rate": 8, + "force_size": "Custom Width", + "custom_width": 304, + "custom_height": 312, + "frame_load_cap": 16, + "skip_first_frames": 1, + "select_every_nth": 1, + "choose video to upload": "image", + "videopreview": { + "hidden": false, + "paused": false, + "params": { + "frame_load_cap": 0, + "skip_first_frames": 0, + "force_rate": 0, + "filename": "leader.webm", + "type": "input", + "format": "video/mp4", + "force_size": "410.4x?", + "select_every_nth": 1 + } + } + } + }, + { + "id": 3, + "type": "VHS_VideoCombine", + "pos": [ + 629, + 222 + ], + "size": { + "0": 315, + "1": 250 + }, + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "name": "images", + "type": "IMAGE", + "link": 1 + } + ], + "outputs": [], + "properties": { + "Node name for S&R": "VHS_VideoCombine" + }, + "widgets_values": { + "frame_rate": 8, + "loop_count": 0, + "filename_prefix": "AnimateDiff", + "format": "video/webm", + "pingpong": false, + "save_image": false, + "crf": 20, + "save_metadata": false, + "audio_file": "", + "videopreview": { + "hidden": false, + "paused": false + } + } + } + ], + "links": [ + [ + 1, + 1, + 0, + 3, + 0, + "IMAGE" + ] + ], + "groups": [], + "config": {}, + "extra": {}, + "version": 0.4, + "tests": { + "3": [{"type": "video", "key": "width", "value": 304}, + {"type": "video", "key": "height", "value": 232}, + {"type": "compare", "filename": "custom_nodes/ComfyUI-VideoHelperSuite/tests/outputs/simple.webm", "tolerance": 0.02} + ], + "length": 1 + } +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/16bit-png.json b/custom_nodes/comfyui-videohelpersuite/video_formats/16bit-png.json new file mode 100644 index 0000000000000000000000000000000000000000..b768bdbcfe8950cb7bde37d02444f0191ad29f51 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/16bit-png.json @@ -0,0 +1,9 @@ +{ + "main_pass": + [ + "-n", + "-pix_fmt", "rgba64" + ], + "input_color_depth": "16bit", + "extension": "%03d.png" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/8bit-png.json b/custom_nodes/comfyui-videohelpersuite/video_formats/8bit-png.json new file mode 100644 index 0000000000000000000000000000000000000000..88b27a75f9c979f18979ec46da8c8843a82e0759 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/8bit-png.json @@ -0,0 +1,7 @@ +{ + "main_pass": + [ + "-n" + ], + "extension": "%03d.png" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/ProRes.json b/custom_nodes/comfyui-videohelpersuite/video_formats/ProRes.json new file mode 100644 index 0000000000000000000000000000000000000000..eddd3acf931c5ab3486197c18ba88a3959d3e1d6 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/ProRes.json @@ -0,0 +1,22 @@ +{ + "main_pass": + [ + "-n", "-c:v", "prores_ks", + "-profile:v", [["$profile"]], + ["profile", { + "lt": [[]], "1": [[]], "standard": [[]], "2": [[]], "hq": [[]], "3": [[]], + "4": ["has_alpha", {"True": [["-pix_fmt", "yuva444p10le"]], + "False": [["-pix_fmt", "yuv444p10le"]]}], + "4444": ["has_alpha", {"True": [["-pix_fmt", "yuva444p10le"]], + "False": [["-pix_fmt", "yuv444p10le"]]}], + "4444xq": ["has_alpha", {"True": [["-pix_fmt", "yuva444p10le"]], + "False": [["-pix_fmt", "yuv444p10le"]]}] + }], + "-vf", "scale=out_color_matrix=bt709", + "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "pcm_s16le"], + "extension": "mov", + "extra_widgets": [["profile", ["lt", "standard", "hq", "4444", "4444xq"], {"default": "hq"}]] +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/av1-webm.json b/custom_nodes/comfyui-videohelpersuite/video_formats/av1-webm.json new file mode 100644 index 0000000000000000000000000000000000000000..659ea166fb140e18109f09f3d5710363af9c2294 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/av1-webm.json @@ -0,0 +1,16 @@ +{ + "main_pass": + [ + "-n", "-c:v", "libsvtav1", + "-pix_fmt", ["pix_fmt", ["yuv420p10le", "yuv420p"]], + "-crf", ["crf","INT", {"default": 23, "min": 0, "max": 100, "step": 1}], + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "libopus"], + "input_color_depth": ["input_color_depth", ["8bit", "16bit"]], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "extension": "webm", + "environment": {"SVT_LOG": "1"} +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/ffmpeg-gif.json b/custom_nodes/comfyui-videohelpersuite/video_formats/ffmpeg-gif.json new file mode 100644 index 0000000000000000000000000000000000000000..3ca0865033c7dabc06176c737652f375e7badb27 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/ffmpeg-gif.json @@ -0,0 +1,8 @@ +{ + "main_pass": + [ + "-n", + "-filter_complex", ["dither", ["bayer", "heckbert", "floyd_steinberg", "sierra2", "sierra2_4a", "sierra3", "burkes", "atkinson", "none"], {"default": "sierra2_4a"}, "[0:v] split [a][b]; [a] palettegen=reserve_transparent=on:transparency_color=ffffff [p]; [b][p] paletteuse=dither=$val"] + ], + "extension": "gif" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/ffv1-mkv.json b/custom_nodes/comfyui-videohelpersuite/video_formats/ffv1-mkv.json new file mode 100644 index 0000000000000000000000000000000000000000..3248e5fc6425feeb1d55ccdb8bd8d49cb0e3401c --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/ffv1-mkv.json @@ -0,0 +1,18 @@ +{ + "main_pass": [ + "-n", + "-c:v", "ffv1", + "-level", ["level", ["0", "1", "3"], {"default": "3"}], + "-coder", ["coder", ["0", "1", "2"], {"default": "1"}], + "-context", ["context", ["0", "1"], {"default": "1"}], + "-g", ["gop_size", "INT", {"default": 1, "min": 1, "max": 300, "step": 1}], + "-slices", ["slices", ["4", "6", "9", "12", "16", "20", "24", "30"], {"default": "16"}], + "-slicecrc", ["slicecrc", ["0", "1"], {"default": "1"}], + "-pix_fmt", ["pix_fmt", ["rgba64le", "bgra", "yuv420p", "yuv422p", "yuv444p", "yuva420p", "yuva422p", "yuva444p", "yuv420p10le", "yuv422p10le", "yuv444p10le", "yuv420p12le", "yuv422p12le", "yuv444p12le", "yuv420p14le", "yuv422p14le", "yuv444p14le", "yuv420p16le", "yuv422p16le", "yuv444p16le", "gray", "gray10le", "gray12le", "gray16le"], {"default": "rgba64le"}] + ], + "audio_pass": ["-c:a", "flac"], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "trim_to_audio": ["trim_to_audio", "BOOLEAN", {"default": false}], + "input_color_depth": "16bit", + "extension": "mkv" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/gifski.json b/custom_nodes/comfyui-videohelpersuite/video_formats/gifski.json new file mode 100644 index 0000000000000000000000000000000000000000..c96b358ab0ae59dbd97c52a699c8ee9825c5d351 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/gifski.json @@ -0,0 +1,12 @@ +{ + "main_pass": + [ + "-pix_fmt", "yuv444p", + "-vf", "scale=out_color_matrix=bt709:out_range=pc", + "-color_range", "pc" + ], + "extension": "gif", + "gifski_pass": [ + "-Q", ["quality","INT", {"default": 90, "min": 1, "max": 100, "step": 1}] + ] +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/h264-mp4.json b/custom_nodes/comfyui-videohelpersuite/video_formats/h264-mp4.json new file mode 100644 index 0000000000000000000000000000000000000000..4aa0bd57c6a4cc65334d82e5cdaa7aadb8372321 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/h264-mp4.json @@ -0,0 +1,15 @@ +{ + "main_pass": + [ + "-n", "-c:v", "libx264", + "-pix_fmt", ["pix_fmt", ["yuv420p", "yuv420p10le"]], + "-crf", ["crf","INT", {"default": 19, "min": 0, "max": 100, "step": 1}], + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "aac", "-movflags", "use_metadata_tags"], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "trim_to_audio": ["trim_to_audio", "BOOLEAN", {"default": false}], + "extension": "mp4" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/h265-mp4.json b/custom_nodes/comfyui-videohelpersuite/video_formats/h265-mp4.json new file mode 100644 index 0000000000000000000000000000000000000000..559c35073edf9ed74baec0eeffc56fde2b0438c5 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/h265-mp4.json @@ -0,0 +1,17 @@ +{ + "main_pass": + [ + "-n", "-c:v", "libx265", + "-vtag", "hvc1", + "-pix_fmt", ["pix_fmt", ["yuv420p10le", "yuv420p"]], + "-crf", ["crf","INT", {"default": 22, "min": 0, "max": 100, "step": 1}], + "-preset", "medium", + "-x265-params", "log-level=quiet", + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "aac", "-movflags", "use_metadata_tags"], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "extension": "mp4" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_av1-mp4.json b/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_av1-mp4.json new file mode 100644 index 0000000000000000000000000000000000000000..ed8294fcf0c9cc1a4780dc6e2454913a969f533b --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_av1-mp4.json @@ -0,0 +1,15 @@ +{ + "main_pass": + [ + "-n", "-c:v", "av1_nvenc", + "-pix_fmt", ["pix_fmt", ["yuv420p", "p010le"]], + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "aac", "-movflags", "use_metadata_tags"], + "bitrate": ["bitrate","INT", {"default": 10, "min": 1, "max": 999, "step": 1 }], + "megabit": ["megabit","BOOLEAN", {"default": true}], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "extension": "mp4" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_h264-mp4.json b/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_h264-mp4.json new file mode 100644 index 0000000000000000000000000000000000000000..f8676304effd31fb41ae13dd972d92736b122009 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_h264-mp4.json @@ -0,0 +1,15 @@ +{ + "main_pass": + [ + "-n", "-c:v", "h264_nvenc", + "-pix_fmt", ["pix_fmt", ["yuv420p", "p010le"]], + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "aac", "-movflags", "use_metadata_tags"], + "bitrate": ["bitrate","INT", {"default": 10, "min": 1, "max": 999, "step": 1 }], + "megabit": ["megabit","BOOLEAN", {"default": true}], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "extension": "mp4" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_hevc-mp4.json b/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_hevc-mp4.json new file mode 100644 index 0000000000000000000000000000000000000000..a374ecde4d0e0c2a2f0042c282f8b1915e307aca --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/nvenc_hevc-mp4.json @@ -0,0 +1,16 @@ +{ + "main_pass": + [ + "-n", "-c:v", "hevc_nvenc", + "-vtag", "hvc1", + "-pix_fmt", ["pix_fmt", ["yuv420p", "p010le"]], + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "aac", "-movflags", "use_metadata_tags"], + "bitrate": ["bitrate","INT", {"default": 10, "min": 1, "max": 999, "step": 1 }], + "megabit": ["megabit","BOOLEAN", {"default": true}], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "extension": "mp4" +} diff --git a/custom_nodes/comfyui-videohelpersuite/video_formats/webm.json b/custom_nodes/comfyui-videohelpersuite/video_formats/webm.json new file mode 100644 index 0000000000000000000000000000000000000000..6323f619ad2850ba60cd5e31b32cdfc060e030c9 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/video_formats/webm.json @@ -0,0 +1,16 @@ +{ + "main_pass": + [ + "-n", + "-pix_fmt", ["pix_fmt",["yuv420p","yuva420p"]], + "-crf", ["crf","INT", {"default": 20, "min": 0, "max": 100, "step": 1}], + "-b:v", "0", + "-vf", "scale=out_color_matrix=bt709", + "-color_range", "tv", "-colorspace", "bt709", "-color_primaries", "bt709", "-color_trc", "bt709" + ], + "fake_trc": "bt709", + "audio_pass": ["-c:a", "libvorbis"], + "save_metadata": ["save_metadata", "BOOLEAN", {"default": true}], + "trim_to_audio": ["trim_to_audio", "BOOLEAN", {"default": false}], + "extension": "webm" +} diff --git a/custom_nodes/comfyui-videohelpersuite/videohelpersuite/batched_nodes.py b/custom_nodes/comfyui-videohelpersuite/videohelpersuite/batched_nodes.py new file mode 100644 index 0000000000000000000000000000000000000000..624bb84f73cce1d8e51abf4fef0b80367ad4b9a3 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/videohelpersuite/batched_nodes.py @@ -0,0 +1,56 @@ +import torch +from nodes import VAEEncode +from comfy.utils import ProgressBar + + +class VAEDecodeBatched: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "samples": ("LATENT", ), + "vae": ("VAE", ), + "per_batch": ("INT", {"default": 16, "min": 1}) + } + } + + CATEGORY = "Video Helper Suite 🎥🅥🅗🅢/batched nodes" + + RETURN_TYPES = ("IMAGE",) + FUNCTION = "decode" + + def decode(self, vae, samples, per_batch): + decoded = [] + pbar = ProgressBar(samples["samples"].shape[0]) + for start_idx in range(0, samples["samples"].shape[0], per_batch): + decoded.append(vae.decode(samples["samples"][start_idx:start_idx+per_batch])) + pbar.update(per_batch) + return (torch.cat(decoded, dim=0), ) + + +class VAEEncodeBatched: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "pixels": ("IMAGE", ), "vae": ("VAE", ), + "per_batch": ("INT", {"default": 16, "min": 1}) + } + } + + CATEGORY = "Video Helper Suite 🎥🅥🅗🅢/batched nodes" + + RETURN_TYPES = ("LATENT",) + FUNCTION = "encode" + + def encode(self, vae, pixels, per_batch): + t = [] + pbar = ProgressBar(pixels.shape[0]) + for start_idx in range(0, pixels.shape[0], per_batch): + try: + sub_pixels = vae.vae_encode_crop_pixels(pixels[start_idx:start_idx+per_batch]) + except: + sub_pixels = VAEEncode.vae_encode_crop_pixels(pixels[start_idx:start_idx+per_batch]) + t.append(vae.encode(sub_pixels[:,:,:,:3])) + pbar.update(per_batch) + return ({"samples": torch.cat(t, dim=0)}, ) diff --git a/custom_nodes/comfyui-videohelpersuite/videohelpersuite/documentation.py b/custom_nodes/comfyui-videohelpersuite/videohelpersuite/documentation.py new file mode 100644 index 0000000000000000000000000000000000000000..857de6c6a5b4a3917f49d0b1c5c142f274203f22 --- /dev/null +++ b/custom_nodes/comfyui-videohelpersuite/videohelpersuite/documentation.py @@ -0,0 +1,616 @@ +from .logger import logger + +def image(src): + return f'' +def video(src): + return f'