Research map / Image generation / Fast diffusion sampling
Budget-Constrained Step-Level Diffusion Caching
Mingkun Lei, Tong Zhao, Liangyu Yuan, Chi 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 |
|---|---|---|---|
| search a static budgeted cache policy instead of threshold-triggered caching, same latency | BudCache vs TeaCache · DrawBench, FLUX.1-dev, 2.51x | LPIPS: 0.3218 -> 0.1759 | Improved |
| use BudCache-ultra (3.00x) versus slower TeaCache (2.51x) | BudCache-ultra · DrawBench, FLUX.1-dev | PSNR: 18.401 -> 22.163 | Improved |
| transfer searched cache mask to other solvers (iPNDM, DPM-Solver++) (ablation) | BudCache-base · DrawBench, FLUX.1-dev | PSNR: 15.35 -> 23.80 (iPNDM vs TeaCache) | Improved |
| transfer searched cache mask to 512x512, 768x1024 and CFG 5/7 (ablation) | BudCache-base · DrawBench, FLUX.1-dev | LPIPS: 0.3154 -> 0.1167 (512x512) | Improved |
| transfer cache policy to GenEval prompts (ablation) | BudCache-base · GenEval, FLUX.1-dev | LPIPS: 0.3349 -> 0.1388 | Improved |
| re-search cache policy on a distilled few-step model | BudCache-base on Hyper-SD (8 steps) · DrawBench | LPIPS: 0.1982 -> 0.1417 | Improved |
| transfer mask searched on Wan2.1 1.3B to 14B (ablation) | BudCache · Wan2.1-T2V-14B, 30 prompts | PSNR: 26.70 -> 27.94 | Improved |
| apply cache-aware schedule alignment on top of the cache policy at 5 NFE (ablation) | BudCache · DrawBench, 5 NFE | ImageReward: 0.7808 -> 0.7888 | Improved |
| apply schedule alignment alone on the original uniform solver (ablation) | Original solver · DrawBench, 5/6/7 NFE | ImageReward: 0.5896 -> 0.5991 (5 NFE); 0.8519 -> 0.8099 (7 NFE) | Mixed |
| use optimized cache policy instead of simply reducing step count (ablation) | BudCache · DrawBench, 5 NFE | ImageReward: 0.5896 -> 0.7808 | Improved |
| increase SA search intensity (R1 SA20 to R2 SA100) (ablation) | BudCache search, 9 NFE · DrawBench | LPIPS: 0.2020 -> 0.2036 | No clear effect |
| increase nominal solver steps K under fixed 10 NFE (ablation) | BudCache · DrawBench, 10 NFE | quality: best at K=30 | Mixed |
Closest papers by meaning
- OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models (ECCV 2026 · similarity 0.93)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · similarity 0.90)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.89)
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.89)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.88)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.88)
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.
