CoolFace
Modelpublic

AEmotionStudio/sam3.1

sourceHugging Faceotherupdated 4mo agoView on Hugging Face
2likes571downloads
README.md69 linesDownload Raw Back to root
1---2license: other3license_name: sam-license4license_link: https://huggingface.co/facebook/sam3.1/blob/main/LICENSE5library_name: sam36tags:7  - segmentation8  - sam39  - sam3.110  - video11  - multiplex12  - facebookresearch13base_model: facebook/sam3.114---15 16# SAM 3.1 — AEmotionStudio Mirror17 18This is an open mirror of [`facebook/sam3.1`](https://huggingface.co/facebook/sam3.1)19hosted by [Æmotion Studio](https://huggingface.co/AEmotionStudio) so that20`ComfyUI-FFMPEGA` and other downstream tools can auto-download the SAM 3.121multiplex tracker without users having to accept the gated terms one-by-one.22 23> ⚠️ **License:** Use of these weights is governed by the upstream SAM License24> (Meta). See [`LICENSE`](LICENSE) — you accept those terms by downloading25> or using the model. This mirror exists solely to make installation26> ergonomic; it is not a relicensing of the model.27 28## What's here29 30| File | Size | Notes |31|---|---|---|32| `sam3.1_multiplex.safetensors` | ~3.5 GB | Preferred. Safer, mmap-friendly format. |33| `sam3.1_multiplex.pt` | ~3.5 GB | Upstream `.pt` format — kept for byte-for-byte parity with `facebook/sam3.1`. |34| `tokenizer.json`, `vocab.json`, `merges.txt`, `*config*.json` | small | BPE / processor files SAM 3.1 needs at load time. |35 36## SAM 3.1 highlights vs SAM 337 38- New **multiplex tracker** — joint multi-object tracking via shared memory.39- ~7× faster on multi-object video (~32 FPS vs. ~16 FPS for 128 objects).40- ~½ the VRAM (8 GB → 4 GB in FP16) on the same scenes.41- Unchanged image-detection model — the gains are entirely in the video path.42 43## Direct use44 45```python46from sam3.model_builder import build_sam3_multiplex_video_model47 48model = build_sam3_multiplex_video_model(49    checkpoint_path="/path/to/sam3.1_multiplex.pt",50    load_from_HF=False,51    device="cuda",52)53```54 55Requires `sam3` from `git+https://github.com/facebookresearch/sam3.git` (a56revision new enough to include `build_sam3_multiplex_video_model`).57 58## ComfyUI-FFMPEGA59 60`ComfyUI-FFMPEGA` auto-downloads from this mirror first and falls back to61`facebook/sam3.1` if the mirror is unreachable. Pick `sam_version="sam3.1"`62on the Effects Builder or Load Video Path node to use it.63 64## Upstream65 66- Code: <https://github.com/facebookresearch/sam3>67- Original weights: <https://huggingface.co/facebook/sam3.1>68- Paper / blog: see the upstream repo.69