Research map / Video / Video generation
Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation
Yuyang You, Yongzhi Li, Jiahui Li, Yadong Mu, Quan Chen, Peng Jiang
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 |
|---|---|---|---|
| add temporal regularization to DMD (ablation) | DMD 1.3B · VBench Dynamic Degree | Dynamic Degree: 72.22 -> 100.00 | Improved |
| add temporal regularization to DMD (ablation) | DMD 1.3B · VBench Instance Preservation | Instance Preservation: 88.88 -> 85.38 | Hurt |
| add naive regression loss on top of temporal regularization (ablation) | DMD+TR 1.3B · VBench Instance Preservation | Instance Preservation: 85.38 -> 83.04 | Hurt |
| add naive regression loss on top of temporal regularization (ablation) | DMD+TR 1.3B · VBench Dynamic Degree | Dynamic Degree: 100.00 -> 78.61 | Hurt |
| replace naive regression loss with adaptive regression loss (ablation) | DMD+TR+RegLoss 1.3B · VBench Instance Preservation | Instance Preservation: 83.04 -> 92.39 | Improved |
| replace naive regression loss with adaptive regression loss (ablation) | DMD+TR+RegLoss 1.3B · VBench Dynamic Degree | Dynamic Degree: 78.61 -> 99.72 | Improved |
| add inference-time frame interpolation (ablation) | DMD+TR+AdaLoss 1.3B · Inference time | time (s): 10.8s -> 7.8s | Improved |
| add inference-time frame interpolation (ablation) | DMD+TR+AdaLoss 1.3B · VBench Instance Preservation | Instance Preservation: 92.39 -> 91.81 | No clear effect |
| add inference-time frame interpolation (ablation) | DMD+TR+AdaLoss 1.3B · VBench Dynamic Degree | Dynamic Degree: 99.72 -> 97.77 | No clear effect |
| use full method instead of DMD baseline | Wan2.1-T2V-1.3B 4-step · VBench2 Total | score: 53.63 -> 55.08 | Improved |
| use full method instead of DMD baseline | Wan2.1-T2V-14B 4-step · VBench2 Total | score: 56.87 -> 59.06 | Improved |
| use full method instead of DMD baseline | Wan2.1-T2V-14B 4-step · VBench1 Total | score: 79.63 -> 82.57 | Improved |
Closest papers by meaning
- When Distillation Breaks Motion Control: Restoring Generative Trajectories for Fast Video Generators (ECCV 2026 · similarity 0.92)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.92)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.91)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.91)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.90)
- Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation (ICLR 2025 · similarity 0.89)
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.
