Research map / Learning theory and algorithms / Submodular and game algorithms
EigenCache: Rethinking Diffusion Acceleration as Covariance-Optimal Forecasting and Submodular Information Allocation
Chenyang Xu, Dezhen Wang, Lin Chen, Kepeng Lin, Hao Wang
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 Hermite/Taylor polynomial forecasting with Kriging covariance predictor | EigenCache vs HiCache · FLUX.1-dev moderate reduction | ImageReward: 0.9979 -> 0.9985 | Improved |
| replace HiCache with EigenCache at aggressive acceleration | HiCache (N=9) · FLUX.1-dev aggressive reduction | ImageReward: 0.9113 -> 0.9120 | No clear effect |
| replace HiCache with EigenCache at aggressive acceleration | HiCache (N=9) · FLUX.1-dev aggressive reduction | LPIPS: 0.4374 -> 0.4150 | Improved |
| replace HiCache with EigenCache at aggressive acceleration | HiCache (N=9) · FLUX.1-dev aggressive reduction | PSNR: 28.647 -> 28.850 | Improved |
| use greedy information scheduling instead of fixed interval (ablation) | EigenCache vs HiCache · FLUX.1-dev N=7 | latency (s): 4.84 -> 4.55 | Improved |
| use greedy scheduling instead of uniform scheduling (ablation) | EigenCache scheduler · target posterior variance 0.01 | anchors needed: 12 -> 8 | Improved |
| replace HiCache with EigenCache | HiCache · HunyuanVideo 3.5x | normalized flickering score: 0.955 -> 0.972 | Improved |
| replace TaylorSeer with EigenCache | TaylorSeer · HunyuanVideo 3.5x | normalized flickering score: 0.915 -> 0.972 | Improved |
| replace DeepCache with EigenCache | DeepCache · SDXL COCO-2017 | FID: 18.45 -> 17.82 | Improved |
| replace ToCa with EigenCache | ToCa · DiT-XL/2 standard schedule | FID-50K: 2.95 -> 2.35 | Improved |
| use OUSAC guidance schedule with EigenCache | EigenCache · DiT-XL/2 | FID-50K: 2.35 -> 2.21 | No clear effect |
| use base-model kernel instead of per-LoRA recalibrated oracle kernel (ablation) | EigenCache · FLUX.1-dev with Anime LoRA | ImageReward: 0.9245 -> 0.9238 | No clear effect |
Closest papers by meaning
- HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration (ICLR 2026 · similarity 0.92)
- Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration (CVPR 2026 · similarity 0.91)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.88)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.88)
- Feature Information Dynamics in Diffusion (NeurIPS 2026 · similarity 0.88)
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.87)
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.
