videogenevalkit/checkpoints

videogenevalkit — checkpoint bundle

All model weights needed by the videogenevalkit toolkit. Organized by the benchmark that consumes each set.

Quickstart

hf download videogenevalkit/checkpoints --repo-type dataset --local-dir ckpts

Then the toolkit reads from ckpts/ automatically (path configurable via env vars).

Layout

t2vcompbench/          # T2V-CompBench upstream-mode CV pipeline (6 files, 4.6 GB)
  groundingdino_swint_ogc.pth       # GD-SwinT-OGC backbone (662 MB)
  sam_vit_h_4b8939.pth              # SAM-H (2.4 GB)
  depth_anything_vitl14.pth         # Depth-Anything V1 (1.3 GB)
  cvo_raft_patch_8.pth              # DOT estimator (21 MB)
  movi_f_raft_patch_4_alpha.pth     # DOT refiner (23 MB)
  movi_f_cotracker2_patch_4_wind_8.pth   # DOT tracker / cotracker2 (195 MB)

worldscore/            # WorldScore metric stack (9 files, ~4 GB)
  sam2.1_hiera_large.pt             # SAM2 (898 MB)
  sam2.1_hiera_base_plus.pt         # SAM2 alternative (324 MB)
  VFIMamba.pkl                      # motion_smoothness backbone (264 MB)
  Tartan-C-T-TSKH-spring540x960-M.pth   # SEA-RAFT (79 MB)
  raft-things.pth                   # classic RAFT (21 MB)
  droid.pth                         # DROID-SLAM (16 MB)
  sac+logos+ava1-l14-linearMSE.pth  # LAION aesthetic predictor (4 MB)
  groundingdino_swint_ogc.pth       # ditto (WorldScore wants its own copy)
  sam_vit_h_4b8939.pth              # ditto

vbench/                # VBench v1 prompt-info registry
  VBench_full_info.json
vbench2/               # VBench-2.0 prompt-info registry
  VBench2_full_info.json

hf-models/             # HuggingFace model mirrors (for offline / China-mirror users)
  liuhaotian/llava-v1.6-34b/        # 65 GB, T2V-CompBench MLLM upstream mode
  Qwen/Qwen2.5-7B-Instruct/         # 15 GB, VBench-2.0 Complex_Plot judge
  lmms-lab/LLaVA-Video-7B-Qwen2/    # 15 GB, VBench-2.0 LLaVA dims
  openai/clip-vit-base-patch16/     # 1.2 GB, WorldScore content_alignment
  LiheYoung/depth_anything_vitl14/  # HF-formatted Depth-Anything V1

Licensing

Each weight redistributes from its upstream release under the source's license (predominantly Apache-2.0 and BSD; LLaVA + Qwen are under research-friendly terms). Cite the upstream papers when reporting numbers that depend on these weights.

See also

• Code: https://github.com/videogenevalkit/videogenevalkit
• Smoke-data: https://huggingface.co/datasets/videogenevalkit/smoke-data