Research map / Video / Video generation
Video Generation Models are General-Purpose Vision Learners
Letian Wang, Chuhan Zhang, Rishabh Kabra, Jasper Uijlings, Steven Waslander, Andrew Zisserman, Joao Carreira, Kaiming He, Mykhaylo Andriluka, Eduard Gabriel Bazavan, Andrei Zanfir, Cristian Sminchisescu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use video diffusion (WAN 2.1) pretraining instead of VideoMAE V2 G (ablation) | GenCeption · Depth (Sintel, KITTI, ETH3D average) | average AbsRel: 0.154 -> 0.093 | Improved |
| use video diffusion (WAN 2.1) pretraining instead of V-JEPA H (ablation) | GenCeption · Depth (Sintel, KITTI, ETH3D average) | average AbsRel: 0.281 -> 0.093 | Improved |
| initialize DiT with more pretrained text-to-video layers (ablation) | GenCeption · Depth | performance | Improved |
| scale backbone from WAN 2.1 1.3B to 14B (1 dataset) (ablation) | GenCeption · Depth average | average AbsRel: 0.122 -> 0.093 | Improved |
| scale backbone from WAN 2.1 1.3B to 14B (4 datasets) (ablation) | GenCeption · Depth average | average AbsRel: 0.094 -> 0.071 | Improved |
| increase training datasets from 1 to 4 (14B) (ablation) | GenCeption · Depth average | average AbsRel: 0.093 -> 0.071 | Improved |
| joint multi-task training instead of specialist (ablation) | GenCeption-L · Sintel depth | AbsRel: 0.130 -> 0.156 | Hurt |
| joint multi-task training instead of specialist (ablation) | GenCeption-L · Hi4D normals | mAE: 10.99 -> 11.47 | Hurt |
| joint multi-task training instead of specialist (ablation) | GenCeption-L · Sintel normals | mAE: 29.7 -> 29.3 | Improved |
| joint multi-task training instead of specialist (ablation) | GenCeption-L · Video matting | MSE: 0.0027 -> 0.0010 | Improved |
| joint multi-task training instead of specialist (ablation) | GenCeption-L · Ref-DAVIS | J&F: 76.4 -> 75.8 | No clear effect |
| joint multi-task training with learnable tokens (ablation) | GenCeption · 3D human keypoints | performance | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Genmo — GENMO: A GENeralist Model for Human MOtion (ICCV 2025) · better on 1 of 1 shared result
- NormalCrafter — NormalCrafter: Learning Temporally Consistent Normals from Video Diffusion Priors (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists (ICLR 2024 · similarity 0.89)
- Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation (ICLR 2024 · similarity 0.88)
- DuoGen: Towards Autonomous Interleaved Multimodal Generation (CVPR 2026 · similarity 0.88)
- GenAI Arena: An Open Evaluation Platform for Generative Models (NeurIPS 2024 · similarity 0.88)
- Video-Bench: Human-Aligned Video Generation Benchmark (CVPR 2025 · similarity 0.88)
- Learning Visual Generative Priors without Text (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
