Research map / Video / Video generation
DisCa: Accelerating Video Diffusion Transformers with Distillation-Compatible Learnable Feature Caching
Chang Zou, Changlin Li, Songtao Liu, Zhao Zhong, Kailin Huang, Linfeng Zhang
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 |
|---|---|---|---|
| restrict the MeanFlow average-velocity interval to [0, R] with R=0.2 (Restricted MeanFlow) | MeanFlow distillation of HunyuanVideo, 20 steps · HunyuanVideo VBench | Semantic score: 66.6 -> 70.4 | Improved |
| restrict the MeanFlow average-velocity interval to [0, R] with R=0.2 (Restricted MeanFlow) | MeanFlow distillation of HunyuanVideo, 10 steps · HunyuanVideo VBench | Semantic score: 60.9 -> 68.2 | Improved |
| use R=0.4 instead of R=0.2 in Restricted MeanFlow (ablation) | Restricted MeanFlow, 10 steps · HunyuanVideo VBench | Semantic score: 68.2 -> 67.6 | No clear effect |
| use Restricted MeanFlow instead of original MeanFlow as the base for the cache predictor (ablation) | DisCa · HunyuanVideo VBench | Semantic score: 65.2 -> 69.3 | Improved |
| add learnable neural predictor instead of training-free caching (ablation) | DisCa on distilled model · HunyuanVideo VBench | Semantic score: 67.3 -> 69.3 | Improved |
| add generative adversarial training to the predictor (ablation) | DisCa predictor · HunyuanVideo VBench | Semantic score: 68.5 -> 69.3 | Improved |
| apply DisCa (learnable caching, N=2) on top of Restricted MeanFlow 20 steps | Restricted MeanFlow 20 steps · HunyuanVideo VBench | Semantic score and latency: 70.4 -> 70.8 semantic; 4.97x -> 7.56x speed | Improved |
| apply training-free caching (TaylorSeer, FORA, TeaCache, PAB, Delta-DiT) to the 10-step distilled model | CFG-distilled HunyuanVideo, 10 steps · HunyuanVideo VBench | Semantic score: 59.0 -> 57.5 (FORA N=6) | Hurt |
| use TaylorSeer forecasting caching on distilled model at high acceleration | CFG-distilled HunyuanVideo, 10 steps · HunyuanVideo VBench | Semantic score: 59.0 -> 63.7 | Improved |
| use DisCa N=4 at 11.8x speed versus 50-step CFG-distilled model | DisCa · HunyuanVideo VBench | Total score: 77.9 -> 78.8 | Improved |
| reduce sampling steps of CFG-distilled model from 50 to 10 without other changes | HunyuanVideo · HunyuanVideo VBench | Semantic score: 66.7 -> 59.0 | Hurt |
| use single-tensor learned cache instead of per-layer caches | DisCa · HunyuanVideo peak VRAM | Peak VRAM: 130.7GB -> 97.64GB | Improved |
Closest papers by meaning
- BWCache: Accelerating Video Diffusion Transformers through Block-Wise Caching (ICLR 2026 · similarity 0.92)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.92)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.91)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.90)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 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.
