Research map / Video / Video generation
MagCache: Fast Video Generation with Magnitude-Aware Cache
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 MagCache, a training-free caching method for video diffusion that skips timesteps using a unified magnitude-ratio law of residuals, calibrated with a single prompt.
Problem
Accelerating video diffusion inference via feature caching, where existing methods (e.g., TeaCache) need extensive prompt-curated calibration and risk overfitting and inconsistent outputs.
Approach
Observes that the per-step average magnitude ratio of successive residuals (v_theta(x_t,t) - x_t) decreases monotonically and is nearly prompt- and model-invariant, and uses its cumulative product as a skip-error estimate in an adaptive cache that reuses residuals until an error budget or max skip length is hit.
Key results
- MagCache-fast achieves 2.68× speedup on Wan 2.1 1.3B (LPIPS 0.1748, PSNR 21.54) versus TeaCache-fast 2.14× (LPIPS 0.2412, PSNR 18.14).
- On Open-Sora 1.2, MagCache-fast reaches 2.10× speedup with latency 21.21s, comparable to TeaCache-slow (1.40×, 31.69s), while MagCache-slow gets PSNR 26.93 vs 23.67.
- On HunyuanVideo, MagCache-fast achieves 2.63× speedup with LPIPS 0.0626 versus TeaCache-fast's 2.26× and LPIPS 0.1971.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MagCache-slow instead of TeaCache-slow | Open-Sora 1.2 · Open-Sora 1.2 (51 frames, 480P) | LPIPS: 0.1303 -> 0.0827 | Helped |
| use MagCache-fast instead of TeaCache-fast | Open-Sora 1.2 · Open-Sora 1.2 (51 frames, 480P) | speedup: 2.05× -> 2.10× | Helped |
| use MagCache-slow instead of TeaCache-slow | Wan 2.1 1.3B · Wan 2.1 1.3B (81 frames, 480P) | LPIPS: 0.1258 -> 0.1206 | Helped |
| use MagCache-fast instead of TeaCache-fast | Wan 2.1 1.3B · Wan 2.1 1.3B (81 frames, 480P) | speedup: 2.14× -> 2.68× | Helped |
| use MagCache-slow instead of TeaCache-slow | HunyuanVideo · HunyuanVideo (129 frames, 540P) | LPIPS: 0.1832 -> 0.0377 | Helped |
| use MagCache-fast instead of TeaCache-fast | HunyuanVideo · HunyuanVideo (129 frames, 540P) | LPIPS: 0.1971 -> 0.0626 | Helped |
| use MagCache instead of TeaCache | CogVideoX 2B · CogVideoX 2B (49 frames, 480P) | LPIPS: 0.1221 -> 0.0787 | Helped |
| use MagCache-slow instead of TeaCache-slow | Flux · Flux Text-to-Image (1024 × 1024) | LPIPS: 0.2687 -> 0.2043 | Helped |
| increase maximum skip length K from 2 to 4 (ablation) | MagCache on Wan 2.1 · Wan 2.1 1.3B | LPIPS: 0.0940 -> 0.1375 | Hurt |
| increase maximum skip length K from 2 to 4 (ablation) | MagCache on Wan 2.1 · Wan 2.1 1.3B | speedup: 2.0× -> 2.4× | Helped |
| decrease error threshold δ from 0.06 to 0.03 (ablation) | MagCache-slow · Open-Sora | LPIPS: 0.1414 -> 0.1389 | Helped |
| calibrate with all 944 prompts rather than one random prompt (ablation) | MagCache on Wan 2.1 · Wan 2.1 1.3B | LPIPS: 0.1206 -> 0.1162 | No clear effect |
Surprising findings
- Calibrating with a single random prompt gives nearly identical results to calibrating with all 944 prompts, and an outlier prompt also performs similarly.
Limitations
- Evaluated only on video (and one image) diffusion models; generalization to other tasks is not verified.
- Error modeling is an approximation, so error can accumulate over long skip sequences, motivating the max skip length K.
- No error bars reported due to resource limits.
Experimental setup
Datasets: VBench prompts
Benchmarks: VBench
Metrics: FLOPs latency speedup LPIPS SSIM PSNR
Baselines: TeaCache FasterCache PAB T-GATE Δ-DiT DuCa TaylorSeer AdaCache
Builds on: TeaCache FasterCache PAB FlashAttention
Closest papers by meaning
- Accelerating Autoregressive Video Diffusion via History-Guided Cache and Residual Correction (CVPR 2026 · similarity 0.92)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.90)
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.90)
- LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation (NeurIPS 2025 · similarity 0.89)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (NeurIPS 2026 · similarity 0.88)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.87)
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.