Research map / Video / Video generation
FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FasterCache, a training-free acceleration method for video diffusion transformers combining dynamic attention-feature reuse with frequency-aware CFG output caching.
Problem
Cache-based acceleration of video diffusion models degrades detail when adjacent-step features are reused directly, and classifier-free guidance doubles compute without acceleration methods targeting it.
Approach
Reuse attention features computed every other step with a first-order bias term (difference of cached features scaled by a linearly increasing weight), and cache CFG unconditional outputs as high- and low-frequency biases relative to conditional outputs, enhanced by timestep-dependent weights during reuse.
Key results
- On Vchitect-2.0, FasterCache achieves 1.67x speedup (260.32s to 156.13s) with VBench 80.84% vs baseline 80.80%.
- On Open-Sora-Plan, 1.68x speedup with VBench 80.19% vs 80.16% baseline, outperforming PAB (1.32x) and Δ-DiT.
- User preference study shows raters favor FasterCache over PAB (69-74%) and Δ-DiT (77-81%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply FasterCache instead of PAB | Open-Sora 1.2 · Open-Sora 1.2 (192 frames, 480P) | speedup: 1.23× -> 1.62× | Helped |
| Apply FasterCache instead of PAB | Open-Sora-Plan · Open-Sora-Plan (65 frames, 512×512) | speedup: 1.32× -> 1.68× | Helped |
| Apply FasterCache instead of PAB | Latte · Latte (16 frames, 512×512) | speedup: 1.28× -> 1.54× | Helped |
| Apply FasterCache instead of PAB | CogVideoX · CogVideoX (48 frames, 480P) | speedup: 1.35× -> 1.62× | Helped |
| Apply FasterCache instead of PAB | Vchitect-2.0 · Vchitect-2.0 (40 frames, 480P) | speedup: 1.26× -> 1.67× | Helped |
| Replace vanilla feature reuse with dynamic feature reuse (ablation) | FasterCache on Open-Sora · Open-Sora (48 frames, 480P) | VBench: 78.34% -> 78.69% | Helped |
| Add dynamic frequency-bias enhancement to CFG-Cache (ablation) | FasterCache on Open-Sora · Open-Sora (48 frames, 480P) | VBench: 78.42% -> 78.69% | Helped |
| Use constant w(t)=0.5 instead of linear weights (ablation) | Dynamic feature reuse on Open-Sora · Open-Sora (48 frames, 480P) | LPIPS: 0.0590 -> 0.0615 | Hurt |
| Use learned weights instead of linear weights (ablation) | Dynamic feature reuse on Open-Sora · Open-Sora (48 frames, 480P) | LPIPS: 0.0596 -> 0.0590 | No clear effect |
| Increase Dynamic FR caching interval from 2 to 3 (ablation) | FasterCache on Open-Sora · Open-Sora (48 frames, 480P) | LPIPS: 0.0590 -> 0.0698 | Hurt |
| Increase Dynamic FR caching interval from 2 to 4 (ablation) | FasterCache on Open-Sora · Open-Sora (48 frames, 480P) | LPIPS: 0.0590 -> 0.0751 | Hurt |
| Increase Dynamic FR caching interval from 2 to 5 (ablation) | FasterCache on Open-Sora · Open-Sora (48 frames, 480P) | LPIPS: 0.0590 -> 0.0897 | Hurt |
Surprising findings
- Directly reusing conditional outputs as unconditional outputs (naive CFG reuse) substantially degrades fine detail, and reusing unconditional outputs from adjacent steps causes error accumulation.
Limitations
- Degrades when base model quality is suboptimal.
- In complex scenes with large motion, may produce degraded results requiring manual hyperparameter tuning.
- Adaptive caching strategies not yet explored.
Experimental setup
Datasets: VBench prompts
Benchmarks: VBench
Metrics: VBench LPIPS PSNR SSIM MACs latency speedup
Baselines: PAB Δ-DiT
Builds on: PAB Δ-DiT FreeU DeepCache classifier-free guidance
Closest papers by meaning
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.90)
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.90)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.90)
- Fast and Memory-Efficient Video Diffusion Using Streamlined Inference (NeurIPS 2024 · similarity 0.90)
- Ca2-VDM: Efficient Autoregressive Video Diffusion Model with Causal Generation and Cache Sharing (ICML 2025 · similarity 0.89)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · 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.