Simplify model card and refresh README audit entry

README.md

@@ -13,101 +13,40 @@ tags:
 
 # EdenLive 14B — Diffusers
 
-This directory packages the 14B EdenLive bidirectional and causal models as standard Diffusers components. Both pipelines generate a complete video from a proxy video, a reference image, and text. Causal generation can change the prompt between chunks within one call.
+[EdenLive](https://github.com/VAST-AI-Research/EdenLive) generates video from a proxy video, a reference image, and a text prompt. This repository includes the 14B bidirectional and causal models for Diffusers.
 
-The source is [VAST-AI/EdenLive](https://huggingface.co/VAST-AI/EdenLive/tree/de40b28f3f6eb6a0f12bb3d3967ca35578273713), revision `de40b28f3f6eb6a0f12bb3d3967ca35578273713`. The 14B causal checkpoint in that release is SGF-v6. Conversion preserves the source tensor precision: transformer weights are saved in FP32, and the examples load them in BF16. The trained DINOv3 encoder and projection are included.
+## Usage
 
-Use a Diffusers checkout containing `EdenLivePipeline` and `EdenLiveCausalPipeline`, plus Transformers with DINOv3 support. This package is published at [`VAST-AI/EdenLive-14B-Diffusers`](https://huggingface.co/VAST-AI/EdenLive-14B-Diffusers).
+Install the EdenLive integration and dependencies:
 
-## Static prompt
+```bash
+pip install "git+https://github.com/skygoo2000/diffusers.git@edenlive" \
+  torch torchvision "transformers>=4.56.0" accelerate sentencepiece ftfy imageio imageio-ffmpeg
+```
 
-The root configuration selects the bidirectional model and the full Wan VAE. Supply proxy frames sampled at 16 fps; the output resolution defaults to 480 × 832.
+Prepare a reference image and a proxy video with at least 49 frames sampled at 16 fps.
 
 ```python
 import torch
 from diffusers import EdenLivePipeline
 from diffusers.utils import export_to_video, load_image, load_video
 
-model_id = "VAST-AI/EdenLive-14B-Diffusers"
-dtype = {"default": torch.bfloat16, "proxy_encoder": torch.float32, "proxy_projection": torch.float32}
-pipe = EdenLivePipeline.from_pretrained(model_id, torch_dtype=dtype).to("cuda")
-image = load_image("reference.png")
-video = load_video("proxy-16fps.mp4")
+pipe = EdenLivePipeline.from_pretrained(
+    "VAST-AI/EdenLive-14B-Diffusers",
+    torch_dtype={"default": torch.bfloat16, "proxy_encoder": torch.float32, "proxy_projection": torch.float32},
+).to("cuda")
+
 frames = pipe(
-    image=image,
-    video=video,
-    prompt="A character walks through snow.",
-    num_frames=85,
-    generator=torch.Generator(device="cuda").manual_seed(42),
+    image=load_image("reference.png"),
+    video=load_video("proxy-16fps.mp4"),
+    prompt="A character explores a snowy landscape.",
+    num_frames=49,
 ).frames[0]
 export_to_video(frames, "edenlive.mp4", fps=16)
 ```
 
-The proxy must contain at least the requested number of frames. Valid output lengths are $4k+1$ frames, where $k$ is a nonnegative integer. For limited GPU memory, load on CPU and call `pipe.enable_model_cpu_offload()` before generation. The Diffusers `examples/edenlive/inference_edenlive.py` example supports context parallelism and FSDP on multiple GPUs.
-
-## Causal and dynamic prompts
-
-Override the transformer and scheduler using their standard component loaders:
-
-```python
-from diffusers import EdenLiveCausalPipeline, EdenLiveFlowMatchEulerDiscreteScheduler, EdenLiveTransformer3DModel
-
-causal_pipe = EdenLiveCausalPipeline.from_pretrained(
-    model_id,
-    transformer=EdenLiveTransformer3DModel.from_pretrained(
-        model_id, subfolder="transformer_causal", torch_dtype=torch.bfloat16
-    ),
-    scheduler=EdenLiveFlowMatchEulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler_causal"),
-    torch_dtype=dtype,
-).to("cuda")
-a = "A character walks through snow."
-b = "The same character walks through a warm forest."
-frames = causal_pipe(
-    image=image,
-    video=video,
-    chunk_prompts=[a, a, a, b, b, b, a, a],
-    num_frames=85,
-    guidance_scale=3.0,
-    negative_prompt="blurry, distorted",
-    generator=torch.Generator(device="cuda").manual_seed(42),
-).frames[0]
-export_to_video(frames, "edenlive-causal.mp4", fps=16)
-```
-
-At the default chunk size of three latent frames, 85 video frames require eight chunk prompts. Each prompt applies to one latent chunk, including the final short chunk. Repeated text is encoded once. `chunk_prompts` is for one video, overrides a single `prompt`, and cannot be combined with positive `prompt_embeds`; the negative prompt stays fixed. For a static causal prompt, replace `chunk_prompts` with `prompt=a`. History belongs to the current call and is cleared when it ends.
-
-## Optional LightTAE decoder
-
-The full Wan VAE is the default decoder. To use LightTAE, load it explicitly and pass it as the `decoder` component when constructing either pipeline:
-
-```python
-from diffusers import EdenLiveLightTAEDecoder
-
-decoder = EdenLiveLightTAEDecoder.from_pretrained(
-    model_id, subfolder="decoder", torch_dtype=torch.bfloat16
-)
-pipe = EdenLivePipeline.from_pretrained(model_id, decoder=decoder, torch_dtype=dtype).to("cuda")
-```
-
-For the causal pipeline, add `decoder=decoder` to its loading call above. `save_pretrained()` saves the selected components; reload that saved directory using the corresponding pipeline class.
-
-## Conversion and verification
-
-```bash
-python scripts/convert_edenlive_to_diffusers.py \
-  --source /path/to/original/EdenLive --size 14b --output ./EdenLive-14B-Diffusers
-```
-
-`conversion_manifest.json` records the names, shapes, dtypes, and SHA256 hashes of the converted tensor values. The converter strictly checks the state dictionaries and verifies saved/reloaded values. `source_manifest.json` records the original source-file hashes and revision. `artifact_manifest.json` records the files in this package. The directory contains all its components and has no internal symlinks.
+See the [usage guide](https://github.com/skygoo2000/diffusers/blob/edenlive/docs/source/en/api/pipelines/edenlive.md) for causal generation, dynamic prompts, and LightTAE, and the [inference examples](https://github.com/skygoo2000/diffusers/tree/edenlive/examples/edenlive#multiple-gpus) for multi-GPU usage.
 
 ## License
 
-The package contains components with different licenses, preserved from the original release:
-
-| Components | License and notices |
-| --- | --- |
-| Transformers, Wan VAE, text components, schedulers, LightTAE | [Apache 2.0](LICENSE) |
-| Trained DINOv3 encoder and projection | [DINOv3 License](proxy_encoder/LICENSE.md), [source notice](proxy_encoder/NOTICE.md) |
-| CLIP image encoder and processor | [MIT](image_encoder/LICENSE.md), [source notice](image_encoder/NOTICE.md) |
-
-DINOv3 is by Meta Platforms, Inc.; the trained encoder and projection are derivatives from Vast AI. The original notice and license are retained beside both components. See the [EdenLive project](https://github.com/VAST-AI-Research/EdenLive) for model details and source acknowledgments.
+EdenLive components use [Apache 2.0](LICENSE). DINOv3 components are covered by the [DINOv3 License](proxy_encoder/LICENSE.md). CLIP components use the [MIT license](image_encoder/LICENSE.md).

artifact_manifest.json

@@ -14,7 +14,7 @@
   ],
   "package": "EdenLive-14B-Diffusers",
   "file_count": 64,
-  "total_bytes": 145736866646,
+  "total_bytes": 145736862630,
   "provenance": {
     "source_repository": "VAST-AI/EdenLive",
     "source_revision": "de40b28f3f6eb6a0f12bb3d3967ca35578273713",
@@ -879,13 +879,13 @@
     },
     {
       "relative_path": "README.md",
-      "size_bytes": 5939,
-      "sha256": "25747aedd2b8f6ea9d23a9a4a1f9e7ef5f12c78245a5ad66c99c8708dd1a810d",
-      "mtime_ns": 1792421016687864599,
+      "size_bytes": 1923,
+      "sha256": "f0ec61b9f70c652d62d0d02e1024556ec76e789d7d2b4eb17064f3c0e5210b56",
+      "mtime_ns": 1792482661327287082,
       "stat_unchanged": true,
       "resolved_within_package": true,
       "symlink": false,
-      "elapsed_seconds": 0.0074418336153030396
+      "elapsed_seconds": 0.005263928323984146
     },
     {
       "relative_path": "conversion_manifest.json",
@@ -1537,6 +1537,24 @@
       },
       "previous_manifest_sha256": "5eeae66758d6dba576fa102cab4a304775a378dec51d558219e70ec1641b1607",
       "verification": "README SHA256/size/mtime and package total refreshed; original weight hashes and original audit facts retained."
+    },
+    {
+      "at": "2026-10-20T07:51:01.330746+00:00",
+      "reason": "Simplify the public model card to overview, installation, basic inference, documentation links, and licenses.",
+      "files": [
+        "README.md"
+      ],
+      "before": {
+        "size_bytes": 5939,
+        "sha256": "25747aedd2b8f6ea9d23a9a4a1f9e7ef5f12c78245a5ad66c99c8708dd1a810d",
+        "mtime_ns": 1792421016687864599
+      },
+      "after": {
+        "size_bytes": 1923,
+        "sha256": "f0ec61b9f70c652d62d0d02e1024556ec76e789d7d2b4eb17064f3c0e5210b56",
+        "mtime_ns": 1792482661327287082
+      },
+      "verification": "Only README bytes and its audit entry changed; original weight entries and audit provenance retained."
     }
   ],
   "readme_publication_update": {