Research map / Video / Video generation
PyramidalWan: On Making Pretrained Video Model Pyramidal for Efficient Inference
Denis Korzhenkov, Adil Karjauv, Animesh Karnewar, Mohsen Ghafoorian, Amirhossein Habibian
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 |
|---|---|---|---|
| make pretrained Wan2.1-1.3B pyramidal by finetuning | Wan2.1-1.3B · VBench | Total score: 82.49 -> 82.83 | No clear effect |
| pyramidal inference schedule 20-20-10 | Wan2.1-1.3B · Computational cost | TFLOPs: 2 x 12592 -> 2 x 2821 | Improved |
| pyramidal 2-2-1 schedule instead of 0-0-2 step-distilled | Wan-DMD · Latency | Latency per pass: 43% speedup | Improved |
| add distillation loss in pyramidal finetuning (ablation) | PyramidalWan · VBench-2.0 | Total score: 54.02 -> 54.93 | Improved |
| add teacher-matching loss L_teach to DMD with pyramidal teacher (ablation) | PyramidalWan-DMD-PT* · VBench-2.0 | Total score: 52.36 -> 51.75 | Hurt |
| add teacher-matching loss L_teach to DMD with pyramidal teacher (ablation) | PyramidalWan-DMD-PT* · VBench | Total score: 82.44 -> 82.72 | Improved |
| use simplified DMD objective instead of full objective (ablation) | PyramidalWan-DMD-PT · VBench-2.0 | Total score: 50.67 -> 51.75 | Improved |
| use simplified DMD objective instead of full objective (ablation) | PyramidalWan-DMD-PT · VBench | Total score: 82.56 -> 82.72 | Improved |
| downsample clean latents in pixel space instead of latent space (ablation) | PyramidalWan · visual quality | qualitative | Improved |
| fine-tune patch-pyramidal PPF for video diffusion (ablation) | Wan-PPF · text-to-video quality | qualitative | Hurt |
| use 1 step instead of 2 steps in Wan-DMD (ablation) | Wan-DMD · VBench | Total score: 83.28 -> 79.45 | Hurt |
| use pyramidal DMD with original teacher instead of pyramidal teacher | PyramidalWan · VBench-2.0 | Total score: 51.75 -> 55.36 | Improved |
Closest papers by meaning
- Pyramidal Flow Matching for Efficient Video Generative Modeling (ICLR 2025 · similarity 0.89)
- FIFO-Diffusion: Generating Infinite Videos from Text without Training (NeurIPS 2024 · similarity 0.86)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.86)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.86)
- TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models (CVPR 2024 · similarity 0.86)
- Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model (NeurIPS 2024 · similarity 0.86)
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.
