DotCheck/muybridge-video-v2

Apache-2.0 video-frame AI-likeness head for DotCheck (chronophotography person Muybridge). This repo includes the live .npz head, model card, license, and notices. It is not the product video assembly; clip eyes+ears seating is private Covenant (DotCheck product path).

Field	Value
Hub id	DotCheck/muybridge-video-v2
Wire id	inhouse-video@2
Label	Muybridge (frames person)
Artifact	siglip2_base_patch16_224_linear_head_video_v2.npz
Backbone	google/siglip2-base-patch16-224 (Apache-2.0)
Profile	video (same tower as Vermeer; distinct head)
Frame contract	VS1 holdout / product VS3 densify: edge-aware frames
Aggregate	bag p90 of frame scores (public claim for this head)
Output (frame)	p ∈ [0,1]
Serve	CPU FastAPI image path with profile=video; Express samples frames

Model description

Frozen SigLIP2-base + video-trained linear head. Not a spatiotemporal video transformer. Soundtrack AI-likeness is a separate open head (Helmholtz / inhouse-audio@2). Product clip scores use private Covenant (Muybridge bag + Helmholtz windows + AU6 fuse). Do not rewrite this card’s bag-p90 table as fused-video metrics.

Files in this repo: README.md, LICENSE, NOTICE, CITATION.cff, and the .npz head file(s) listed above.

Architecture

video container
  → sample 3 frames (VS1 timestamps)
  → per-frame: resize ≤256 → SigLIP2 224 → frozen emb → video head
  → bag aggregate: p90({p_frame})

Missing video head at serve → profile=video returns 503 (no silent image-head fallback).

Inference

Open weights: the live .npz head(s) in this repo (Apache-2.0), for use with the frozen upstream backbone named above. This is not a transformers AutoModel.from_pretrained("DotCheck/…") package.

Product scoring: Check or Pro API (below). Leviathan (shared memory and related product path) is not in these files.

curl -sS -X POST "https://dotcheck-server-c221c1f32c68.herokuapp.com/analyze-video" \
  -H "Authorization: Bearer dc_YOUR_KEY" \
  -F "file=@clip.mp4"

UI: https://dotcheck.ai/check · API: https://dotcheck.ai/api · docs: https://dotcheck.ai/docs

Training data

Split	Content
Fit AI	CogVideoX-2b + Wan2.1 (Apache-2.0), VS1 3-frame bags
Holdout AI	Wan2.2-TI2V-5B Diffusers only (FIT_MIX_HOLDOUT=0)
Real	Commons VS1 ffmpeg + ~20% thumb surface + still mix (STILL_FRAC=0.20)

Evidence: video_gates_video_v2.json · VIDEO_GATES_OK. Prior @1 bag bal 0.716 on same bags (archived).

Evaluation

Bag p90 (public claim SSOT)

Metric	Target	Measured
mean P(AI) | real	≤ 0.12	0.004
mean P(AI) | AI	≥ 0.85	0.940
bal_acc @ thr	≥ 0.90	0.985
vs @1 bal (same bags)	≥ baseline	0.985 ≥ 0.716

Secondary single-frame floors (not public headline)

Metric	Target	Measured
mean P(AI) | real	≤ 0.15	0.004
mean P(AI) | AI	≥ 0.80	0.906
separation	≥ 0.45	0.902
bal_acc	≥ 0.88	0.955

Intended use

• Frame-bag AI-likeness for short clips in DotCheck inference.
• Citation of bag-p90 table.

Out of scope

• Product scoring SLA / Leviathan / Covenant seating recipe / FUP via Hub download
• Full temporal models / standalone audio deepfake courtroom forensics (see Helmholtz for audio head)
• Publishing Covenant as Apache weights (assembly stays with DotCheck)
• Legal determinations

Limitations

• Aggregation is unordered frame p90; no explicit motion model.
• Codec / compression / unseen generators can shift scores.
• Silent clips: frames only. When a soundtrack is present, product Covenant fuse with Helmholtz is separate from this artifact’s bag-p90 table (see eval/COVENANT.md / bake-off JSON for operator protocol; fused public floats only after a commercial panel).

License

LICENSE — Apache License 2.0 for DotCheck heads in this repo. Upstream backbones: see NOTICE.

Citation

CITATION.cff · wire inhouse-video@2 / Muybridge@2 · https://dotcheck.ai/docs