Research map / Image generation / Fast diffusion sampling
Evolutionary Caching to Accelerate Your Off-the-Shelf Diffusion Model
Anirud Aggarwal, Abhinav Shrivastava, Matthew Gwilliam
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 |
|---|---|---|---|
| use 100 calibration prompts instead of 33 (ablation) | ECAD PixArt-alpha · Calibration Image Reward | Image Reward: 0.85 -> 0.94 | Improved |
| use population size 72 instead of 24 (ablation) | ECAD PixArt-alpha · Latency of highest-TMACs schedule | ms / img: 122.96 -> 100.68 | Improved |
| generate 10 images per prompt instead of 3 (ablation) | ECAD PixArt-alpha · Calibration Image Reward | Image Reward: 0.96 -> 0.94 | No clear effect |
| allow direct copies of parents instead of crossover probability 1.0 (ablation) | ECAD PixArt-alpha · MJHQ-30K | FID: 9.64 -> 8.60 | Improved |
| use 4-point crossover instead of 1-point crossover (ablation) | ECAD PixArt-alpha · Latency | ms / img: 114.78 -> 100.30 | Improved |
| use mutation probability 5% instead of 1% (ablation) | ECAD PixArt-alpha · MJHQ-30K | FID: 9.11 -> 8.60 | Improved |
| use mutation probability 5% instead of 15% (ablation) | ECAD PixArt-alpha · Latency | ms / img: 127.83 -> 100.30 | Improved |
| calibrate on ChatGPT-generated prompts instead of Image Reward Benchmark (ablation) | ECAD PixArt-alpha · MJHQ-30K | FID: 8.18 -> 8.05 | No clear effect |
| calibrate on 100 painted-landscape prompts instead of Image Reward Benchmark (ablation) | ECAD PixArt-alpha · MJHQ-30K | FID: 8.18 -> 8.55 | No clear effect |
| calibrate on 5-word prompts instead of Image Reward Benchmark (ablation) | ECAD PixArt-alpha · MJHQ-30K | FID: 8.18 -> 7.42 | No clear effect |
| calibrate on 10 prompts instead of 100 (ablation) | ECAD PixArt-alpha · MJHQ-30K | FID: 8.18 -> 10.02 | Hurt |
| run 50 more generations on PixArt-Sigma after transfer from PixArt-alpha (ablation) | ECAD · PixArt-Sigma speedup | speedup: 1.62x -> 1.98x | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiCache — DiCache: Let Diffusion Model Determine Its Own Cache (ICLR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.93)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · similarity 0.91)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.91)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.91)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
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.
