Research map / Image generation / Fast diffusion sampling
HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration
Liang Feng, Shikang Zheng, Jiacheng Liu, Yuqi Lin, Qinming Zhou, Peiliang Cai, Xinyu Wang, Junjie Chen, Chang Zou, Yue Ma, Linfeng 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 |
|---|---|---|---|
| replace Taylor basis with scaled Hermite basis | TaylorSeer on FLUX.1-dev · text-to-image N=7, O=2 | ImageReward: 0.9572 -> 0.9979 | Improved |
| replace Taylor predictor with Hermite predictor in ClusCa | ClusCa on FLUX.1-dev · N=7 | ImageReward: 0.9480 -> 0.9840 | Improved |
| replace Taylor predictor with Hermite predictor in ClusCa | ClusCa on FLUX.1-dev · N=9 | ImageReward: 0.8440 -> 0.8860 | Improved |
| replace Taylor basis with Hermite basis | TaylorSeer on HunyuanVideo · VBench N=7, O=2 | VBench (%): 79.28 -> 79.65 | Improved |
| replace Taylor basis with Hermite basis | TaylorSeer on HunyuanVideo · VBench N=6, O=1 | VBench (%): 79.78 -> 79.89 | Improved |
| replace Taylor basis with Hermite basis | TaylorSeer on DiT-XL/2 · ImageNet ~7.1x | FID: 4.58 -> 4.31 | Improved |
| add contraction factor sigma=0.5 to Hermite basis (vs sigma=1.0) (ablation) | HiCache · FLUX.1-dev 5.55x | ImageReward: 0.7586 -> 0.9979 | Improved |
| vary contraction sigma among 0.4, 0.5, 0.7 (ablation) | HiCache · FLUX.1-dev 5.55x | ImageReward: 0.9683 / 0.9979 / 0.9623 | Mixed |
| apply dual scaling alone to TaylorSeer (Hi-Taylor) (ablation) | TaylorSeer · FLUX.1-dev 5.55x | ImageReward: 0.9572 -> 0.9624 | Improved |
| add Hermite basis on top of dual scaling (HiCache vs Hi-Taylor) (ablation) | HiCache · FLUX.1-dev 5.55x | ImageReward: 0.9624 -> 0.9979 | Improved |
| accelerate with HiCache on image super-resolution | Inf-DiT · NTIRE 2025 protocol, ~5.93x theoretical speedup | PSNR: 30.87 -> 30.18 | Hurt |
Closest papers by meaning
- Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models (CVPR 2026 · similarity 0.92)
- EigenCache: Rethinking Diffusion Acceleration as Covariance-Optimal Forecasting and Submodular Information Allocation (ICML 2026 · similarity 0.92)
- Forecast the Principal, Stabilize the Residual: Subspace-Aware Feature Caching for Diffusion Transformers (CVPR 2026 · similarity 0.92)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.90)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.90)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 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.
