Research map / Video / Video generation
Pusa V1.0: Unlocking Temporal Control in Pretrained Video Diffusion Models via Vectorized Timestep Adaptation
Yaofang Liu, Yumeng REN, Aitor Artola, Yuxuan Hu, Xiaodong Cun, Xiaotong Zhao, Alan Zhao, Raymond H. Chan, Suiyun Zhang, Rui Liu, Dandan Tu, Jean-michel Morel
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 |
|---|---|---|---|
| replace Wan-I2V fine-tuning with vectorized timestep adaptation (Pusa) | Wan2.1-T2V-14B · VBench-I2V | Total score (%): 86.86 -> 87.32 | Improved |
| replace Wan-I2V fine-tuning with Pusa | Wan2.1-T2V-14B · VBench-I2V | Quality score (%): 80.82 -> 79.80 | Hurt |
| replace Wan-I2V training with VTA at equal 900-step budget (480p) (ablation) | Wan2.1-T2V-14B · VBench-I2V subset | Total score (%): 73.21 -> 87.09 | Improved |
| use fully random per-frame timesteps instead of I2V-style fixed first frame (ablation) | Pusa · VBench-I2V subset | Total score (%): 73.27 -> 87.69 | Improved |
| use fully random timesteps instead of PTSS p=0.2 (ablation) | Pusa · VBench-I2V subset | Total score (%): 84.74 -> 87.69 | Improved |
| use fully random timesteps instead of PTSS p=0.8 (ablation) | Pusa · VBench-I2V subset | Total score (%): 86.49 -> 87.69 | Improved |
| increase LoRA rank from 256 to 512 (ablation) | Pusa · VBench-I2V subset | Total score (%): 83.12 -> 86.46 | Improved |
| increase inference steps from 10 to 20 (ablation) | Pusa · VBench-I2V subset | Total score (%): 87.69 -> 87.84 | No clear effect |
| increase inference steps from 2 to 10 (ablation) | Pusa · VBench-I2V subset | Total score (%): 79.92 -> 87.69 | Improved |
| train beyond 900 iterations (ablation) | Pusa LoRA · VBench-I2V subset | Total score (%): 87.69 -> 87.32 | No clear effect |
| swap base model Wan2.1 for Wan2.2 (ablation) | Pusa · VBench-I2V subset | Total score (%): 87.69 -> 87.69 | No clear effect |
| use full fine-tuning instead of LoRA (ablation) | Pusa · VBench-I2V subset | Total score (%): 84.06 vs 87.09 | No clear effect |
Closest papers by meaning
- Ctrl-Adapter: An Efficient and Versatile Framework for Adapting Diverse Controls to Any Diffusion Model (ICLR 2025 · similarity 0.89)
- VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis (ICLR 2025 · similarity 0.89)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.89)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models (CVPR 2024 · similarity 0.88)
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.
