Research map / Image generation / Fast diffusion sampling
OmniCache: A Trajectory-Oriented Global Perspective on Training-Free Cache Reuse for Diffusion Transformer Models
Huanpeng Chu, Wei Wu, Guanyu Feng, Yutao 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 |
|---|---|---|---|
| add noise correction module (ablation) | OmniCache · CogVideoX-5b-I2V-distill I2V | Q-Align: 0.778 -> 0.788 | Improved |
| add noise correction module (ablation) | OmniCache · CogVideoX-5b-I2V-distill I2V | Aesthetic Quality: 0.58 -> 0.593 | Improved |
| add high-pass and low-pass filtering to noise estimate (ablation) | OmniCache · CogVideoX-5b-I2V-distill I2V | Aesthetic Quality: 0.593 -> 0.621 | Improved |
| add high-pass and low-pass filtering to noise estimate (ablation) | OmniCache · CogVideoX-5b-I2V-distill I2V | Q-Align: 0.788 -> 0.792 | Improved |
| cache reuse alone on distilled model (ablation) | OmniCache · CogVideoX-5b-I2V-distill I2V | Aesthetic Quality: 0.59 -> 0.58 | Hurt |
| use OmniCache instead of AdaCache | AdaCache · CogVideoX-5b-I2V-distill I2V | Q-Align: 0.77 -> 0.792 | Improved |
| use OmniCache-slow instead of PAB-slow | PAB-slow · Open-Sora VBench | VBench (%): 78.51 -> 78.83 | Improved |
| use OmniCache-fast instead of PAB-fast | PAB-fast · Latte VBench | VBench (%): 73.13 -> 77.09 | Improved |
| use OmniCache-fast instead of TeaCache-fast | TeaCache-fast · Latte VBench | VBench (%): 76.69 -> 77.09 | Mixed |
| add MoReg to AdaCache-fast (ablation) | AdaCache-fast · Latte VBench | VBench (%): 76.26 -> 76.47 | Mixed |
| use OmniCache-slow on Open-Sora | Open-Sora · Open-Sora VBench | VBench (%): 79.22 -> 78.83 | No clear effect |
| apply OmniCache-fast to DiT-XL/2 class-conditional | DiT-XL/2 · ImageNet 256 | FID: 2.68 -> 2.69 | No clear effect |
Closest papers by meaning
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.93)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.92)
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.92)
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.92)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.92)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · similarity 0.91)
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.
