pointcut-models / README.md
adsfadfadsfadsf's picture
MANIFEST.txt and README for the full mirror
ae073a3 verified
|
Raw History Blame Contribute Delete
3.7 kB
---
license: other
license_name: mixed-open-licenses
license_link: https://huggingface.co/adsfadfadsfadsf/pointcut-models/blob/main/README.md
tags:
- pointcut
- whisper
- lipsync
- musetalk
- tts
- mlx
---
# PointCut model weights
Byte-identical copies of the open model weights the PointCut video editor downloads on a Mac, kept here so that an
install never depends on a file staying up at its author's. PointCut downloads from this repo first (a pinned commit)
and falls back to the original file when a copy can't be had. `MANIFEST.txt` lists every file: `size sha256 path`.
Layout: `<feature id>/<path inside the feature's folder on the Mac>`. Each feature's folder also holds the licence or
model card of every model in it (`LICENSE-*`, `MODELCARD-*`).
| Folder | What it is | Original |
|---|---|---|
| `whisper/` | Whisper large-v3-turbo, 5-bit, whisper.cpp format (MIT) | [ggerganov/whisper.cpp](https://huggingface.co/ggerganov/whisper.cpp) |
| `lipsync/` | MuseTalk 1.5 (MIT): the UNet in half precision (`unet.fp16.safetensors`, converted from `musetalkV15/unet.pth` with `tensor.to(torch.float16)`, nothing else changed — 1.7 GB instead of 3.4 GB), its config and code; SD VAE ft-MSE (MIT); whisper-tiny (Apache-2.0); face-parsing BiSeNet (MIT) with its ResNet-18 backbone (torchvision, BSD-3) | [TMElyralab/MuseTalk](https://huggingface.co/TMElyralab/MuseTalk), [stabilityai/sd-vae-ft-mse](https://huggingface.co/stabilityai/sd-vae-ft-mse), [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny), [ManyOtherFunctions/face-parse-bisent](https://huggingface.co/ManyOtherFunctions/face-parse-bisent), download.pytorch.org |
| `voice/` | Qwen3-TTS 12Hz 1.7B Base, 8-bit MLX (Apache-2.0), in the Hugging Face cache layout mlx-audio loads offline | [mlx-community/Qwen3-TTS-12Hz-1.7B-Base-8bit](https://huggingface.co/mlx-community/Qwen3-TTS-12Hz-1.7B-Base-8bit) |
| `image/` | Z-Image-Turbo, 4-bit MLX weights for mflux (Apache-2.0) | [mflux-community/z-image-turbo-mflux-q4](https://huggingface.co/mflux-community/z-image-turbo-mflux-q4) |
| `music/` | ACE-Step 1.5 turbo with its VAE (MIT) and Qwen3-Embedding-0.6B (Apache-2.0) | [ACE-Step/Ace-Step1.5](https://huggingface.co/ACE-Step/Ace-Step1.5) |
| `faceswap/` | HyperSwap 1b 256 (OpenRAIL-AS, FaceFusion), ArcFace w600k r50 (InsightFace, non-commercial research), GFPGAN 1.4 (Apache-2.0), BiSeNet ResNet-34 face parser (MIT) — ONNX as FaceFusion publishes them; Retouch uses GFPGAN and BiSeNet from here | [facefusion/models-3.3.0](https://huggingface.co/facefusion/models-3.3.0), [facefusion/models-3.0.0](https://huggingface.co/facefusion/models-3.0.0) |
| `cleanup/` | SAM 2.1 Hiera tiny, ONNX (Apache-2.0); LaMa big-lama, ONNX (Apache-2.0) | [onnx-community/sam2.1-hiera-tiny-ONNX](https://huggingface.co/onnx-community/sam2.1-hiera-tiny-ONNX), [Carve/LaMa-ONNX](https://huggingface.co/Carve/LaMa-ONNX) |
| `denoise/` | DeepFilterNet 3: the `deep-filter` program for Apple silicon and the model (MIT / Apache-2.0), release v0.5.6 | [Rikorose/DeepFilterNet](https://github.com/Rikorose/DeepFilterNet) |
| `separate/` | Demucs v4 htdemucs weights and the model code at v4.0.1 (MIT) | [facebookresearch/demucs](https://github.com/facebookresearch/demucs), dl.fbaipublicfiles.com |
| `voicechange/` | OpenVoice v2 tone colour converter and its model code (MIT) | [myshell-ai/OpenVoiceV2](https://huggingface.co/myshell-ai/OpenVoiceV2), [myshell-ai/OpenVoice](https://github.com/myshell-ai/OpenVoice) |
Each model keeps its own licence (see the files next to it); nothing here is relicensed. The repo is a mirror for
one application's installer, not a redistribution of the models under new terms.