Research map / Image generation / Fast diffusion sampling
IdealCache: Rethinking Cache Scheduling in Diffusion Transformers via Ideal Trajectories
shang xue, Li Zhu, Ping Chen
Read the paperCheck your idea against 137,217 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 IdealCache, a feature-caching scheme for Diffusion Transformers that targets an ideal trajectory rather than the standard reference denoising trajectory to schedule cache reuse.
Problem
Feature caching speeds up Diffusion Transformers by reusing computations across timesteps, but existing methods approximate the standard reference denoising trajectory, which caps output quality at that trajectory's quality.
Approach
Argues the standard reference trajectory is an arbitrary optimization target and instead schedules caching to approximate an "ideal trajectory," removing the implicit quality ceiling inherited from the reference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ideal-trajectory schedule (IdealCache B=17) instead of HyCa N=3 at similar speed | Qwen-Image caching · DrawBench | SSIM: 0.763 -> 0.830 | Improved |
| use IdealCache B=17 instead of HyCa N=3 | Qwen-Image caching · DrawBench | ImageReward: 1.1914 -> 1.1767 | Hurt |
| use IdealCache B=7 instead of HyCa N=8 at high compression | Qwen-Image caching · DrawBench | ImageReward: 1.0419 -> 1.1068 | Improved |
| use IdealCache B=9 instead of HyCa N=6 on image editing | Qwen-Image-Edit caching · GEdit-EN Full | Overall Score: 7.45 -> 7.62 | Improved |
| IdealCache B=9 versus unaccelerated 50-step pipeline | Qwen-Image-Edit · GEdit-EN Full | Overall Score: 7.54 -> 7.62 | Improved |
| set source image as ideal target for editing schedule (B=25) | IdealCache on Qwen-Image-Edit · FlowEdit | Background Preservation (SSIM): 0.665 -> 0.687 | Improved |
| IdealCache B=10 versus vanilla 50-step on FlowEdit | Qwen-Image-Edit · FlowEdit | ImageReward: 1.054 -> 0.991 | Hurt |
| IdealCache B=7 versus vanilla 50-step on text-to-image | Qwen-Image · DrawBench | ImageReward: 1.2092 -> 1.1068 | Hurt |
| build schedule graph from one prompt (R=1) instead of R=1000 ensemble (ablation) | IdealCache · DrawBench ImageReward (B=9) | ImageReward: 1.181 -> 1.168 | No clear effect |
| build schedule graph from one prompt (R=1) on GEdit-EN (ablation) | IdealCache · GEdit-EN (B=9) | Overall Score: 7.62 -> 7.58 | No clear effect |
| densify candidate grid from M=50 to M=100 (ablation) | IdealCache · DrawBench 1664x928 (B=10) | SSIM: 0.75 -> 0.78 | Improved |
| densify candidate grid from M=50 to M=100 (ablation) | IdealCache · DrawBench 1664x928 (B=10) | LPIPS: 0.33 -> 0.27 | Improved |
Closest papers by meaning
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.90)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.90)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · similarity 0.90)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.90)
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.90)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · 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.
