Research map / Video / Video generation
LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation
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.
Training-free cache scheduling for diffusion video generation that formulates cache-path selection as a DAG with globally error-weighted edges and solves it with lexicographic minimax path optimization.
Problem
Caching in diffusion-based video generation reduces inference cost but local greedy thresholds ignore temporal heterogeneity and accumulated global error, causing content deviation and detail loss between accelerated and original videos.
Approach
Replace local threshold-based cache decisions with an offline static DAG whose edges are weighted by the final-output error of caching a segment (averaged over prompts and seeds), then pick the path with exactly B full-compute steps that lexicographically minimizes the sorted worst-case edge errors.
Key results
- On Open-Sora 1.2, LeMiCa-slow reaches LPIPS 0.050 and PSNR 31.32 at 1.52x speedup, versus TeaCache-slow LPIPS 0.134 at 1.50x.
- On Latte, LeMiCa-fast achieves 2.93x speedup (3.81 s latency) versus TeaCache-fast 2.60x, with VBench 76.75% vs 76.09%.
- On Qwen-Image, LeMiCa (B=17) reaches LPIPS 0.100 and PSNR 34.326 at 2.45x speedup, versus Cache-DiT (F4W4R0.32) LPIPS 0.137 at 2.42x.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LeMiCa-slow instead of TeaCache-slow | Open-Sora 1.2 · Open-Sora 1.2 (51 frames, 480P) | LPIPS: 0.134 -> 0.050 | Helped |
| Use LeMiCa-fast instead of TeaCache-fast | Latte · Latte (16 frames, 512×512) | speedup: 2.60× -> 2.93× | Helped |
| Use LeMiCa-slow instead of TeaCache-slow | Latte · Latte (16 frames, 512×512) | LPIPS: 0.195 -> 0.091 | Helped |
| Use LeMiCa-slow instead of TeaCache-slow | CogVideoX-2B · CogVideoX (49 frames, 480P) | LPIPS: 0.053 -> 0.023 | Helped |
| Use LeMiCa-fast instead of TeaCache-fast | Open-Sora 1.2 · Open-Sora 1.2 (51 frames, 480P) | LPIPS: 0.252 -> 0.187 | Helped |
| Use MiniMax Path instead of Shortest Path (ablation) | LeMiCa · Video reconstruction | VBench: 76.04 -> 79.27 | Helped |
| Use MiniMax Path instead of Shortest Path (ablation) | LeMiCa · Video reconstruction | LPIPS: 0.203 -> 0.143 | Helped |
| Increase DAG construction samples from 1 to 20 (ablation) | LeMiCa · VBench prompts | PSNR: 24.51 -> 24.60 | Helped |
| Increase DAG construction samples from 1 to 20 (ablation) | LeMiCa · VBench prompts | LPIPS: 0.164 -> 0.152 | Helped |
| Use LeMiCa instead of TeaCache | Video generation on IP-VBench · IP-VBench overall | LPIPS: 0.120 -> 0.042 | Helped |
| Use LeMiCa (B=17) instead of Cache-DiT (F4W4R0.32) | Qwen-Image · Qwen-Image | LPIPS: 0.137 -> 0.100 | Helped |
Surprising findings
- A shortest-path (additive) cache selection performed much worse than the minimax criterion (VBench 76.04 vs 79.27, LPIPS 0.203 vs 0.143 in ablation).
- Early-step cache segments induce larger downstream error than later segments of similar length, so segment position matters more than length.
Limitations
- Quality degrades under high acceleration ratios due to fewer forward steps.
- Performance depends on the representational capacity of the underlying diffusion model, struggling when original video quality is low or motion is complex.
- Static offline graph is built from a limited set of prompts and seeds and requires offline construction per model configuration.
Experimental setup
Datasets: T2V-CompBench VBench IP-VBench
Benchmarks: VBench IP-VBench T2V-CompBench
Metrics: VBench LPIPS SSIM PSNR FLOPs latency speedup Image Reward CLIP Score
Baselines: Delta-DiT T-GATE PAB TeaCache Cache-DiT ShortDF
Builds on: TeaCache PAB T-GATE Delta-DiT
Closest papers by meaning
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.89)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.88)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.88)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.87)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · 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.