Modelpublic

AEmotionStudio/sam3.1

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

2likes519downloads

Model Card

SAM 3.1 — AEmotionStudio Mirror

This is an open mirror of `facebook/sam3.1` hosted by Æmotion Studio so that ComfyUI-FFMPEGA and other downstream tools can auto-download the SAM 3.1 multiplex tracker without users having to accept the gated terms one-by-one.

⚠️ License: Use of these weights is governed by the upstream SAM License (Meta). See `LICENSE` — you accept those terms by downloading or using the model. This mirror exists solely to make installation ergonomic; it is not a relicensing of the model.

What's here

File	Size	Notes
`sam3.1_multiplex.safetensors`	~3.5 GB	Preferred. Safer, mmap-friendly format.
`sam3.1_multiplex.pt`	~3.5 GB	Upstream `.pt` format — kept for byte-for-byte parity with `facebook/sam3.1`.
`tokenizer.json`, `vocab.json`, `merges.txt`, `config.json`	small	BPE / processor files SAM 3.1 needs at load time.

SAM 3.1 highlights vs SAM 3

—New multiplex tracker — joint multi-object tracking via shared memory.
—~7× faster on multi-object video (~32 FPS vs. ~16 FPS for 128 objects).
—~½ the VRAM (8 GB → 4 GB in FP16) on the same scenes.
—Unchanged image-detection model — the gains are entirely in the video path.

Direct use

python

from sam3.model_builder import build_sam3_multiplex_video_model

model = build_sam3_multiplex_video_model(
    checkpoint_path="/path/to/sam3.1_multiplex.pt",
    load_from_HF=False,
    device="cuda",
)

Requires sam3 from git+https://github.com/facebookresearch/sam3.git (a revision new enough to include build_sam3_multiplex_video_model).

ComfyUI-FFMPEGA

ComfyUI-FFMPEGA auto-downloads from this mirror first and falls back to facebook/sam3.1 if the mirror is unreachable. Pick sam_version="sam3.1" on the Effects Builder or Load Video Path node to use it.

Upstream

—Code: <https://github.com/facebookresearch/sam3>
—Original weights: <https://huggingface.co/facebook/sam3.1>
—Paper / blog: see the upstream repo.