README.md

---
license: other
language: en
tags: [video-generation, evaluation, checkpoints]
---

# videogenevalkit — checkpoint bundle

Pretrained model weights needed by the
[videogenevalkit](https://github.com/videogenevalkit/videogenevalkit) toolkit
for the **T2V-CompBench** and **WorldScore** adapters in upstream-mode.

We bundle **only** weights that are awkward to fetch from their original
sources (Drive links, expired buckets, etc.). Large models that are already
on HuggingFace are NOT rehosted — for those, the toolkit pulls directly:

- `liuhaotian/llava-v1.6-34b` (68 GB; T2V-CompBench MLLM upstream-mode)
- `lmms-lab/LLaVA-Video-7B-Qwen2` (15 GB; VBench-2.0 LLaVA dims)
- `Qwen/Qwen2.5-7B-Instruct` (15 GB; VBench-2.0 complex_plot judge)

## Contents

| File | Size | Used by | Upstream source |
|---|---:|---|---|
| `groundingdino_swint_ogc.pth` | 662 MB | T2V-CompBench (numeracy, spatial, motion_binding), WorldScore (object_control, motion_accuracy) | [IDEA-Research/GroundingDINO](https://github.com/IDEA-Research/GroundingDINO) — Apache-2.0 |
| `sam_vit_h_4b8939.pth` | 2.4 GB | T2V-CompBench (spatial, motion_binding), WorldScore (motion_accuracy) | [facebookresearch/segment-anything](https://github.com/facebookresearch/segment-anything) — Apache-2.0 |
| `depth_anything_vitl14.pth` | 1.3 GB | T2V-CompBench (spatial_relationships 3D depth) | [LiheYoung/Depth-Anything](https://github.com/LiheYoung/Depth-Anything) — Apache-2.0 |
| `cvo_raft_patch_8.pth` | 21 MB | T2V-CompBench motion_binding (DOT estimator) | [facebookresearch/dot](https://github.com/facebookresearch/dot) — BSD |
| `movi_f_raft_patch_4_alpha.pth` | 23 MB | T2V-CompBench motion_binding (DOT refiner) | [facebookresearch/dot](https://github.com/facebookresearch/dot) — BSD |
| `movi_f_cotracker2_patch_4_wind_8.pth` | 195 MB | T2V-CompBench motion_binding (DOT tracker; cotracker2 backbone) | [facebookresearch/dot](https://github.com/facebookresearch/dot) — BSD |

**Total: ~4.6 GB.**

## Usage

```bash
# After installing videogenevalkit:
videogenevalkit fetch-checkpoints --bench t2vcompbench
# Equivalent to:
hf download videogenevalkit/checkpoints --repo-type dataset \
  --local-dir validation/upstream_ckpts
```

## Licensing

All weights are redistributions of upstream releases under the licenses linked
above (predominantly Apache-2.0 and BSD). Each underlying weight remains under
its source repo's license; cite the upstream papers when reporting numbers
that depend on these weights. The bundling itself imposes no additional terms.

## See also

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