AI Paper Atlas

Research map / Image generation / Editing and style transfer

CacheEdit: Efficient Multi-round Image Editing via Adaptive Token-wise Reuse.

Jinxin Yu, Xueqing Chen, Yudong Pan, Lian Liu, shengwen Liang, Huawei Li, Xiaowei Li, Ying Wang

ICML 2026 · Image generation › Editing and style transfer · approach: Instruction-based diffusion editing

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 changedWhereMeasuredEffect
add CacheEdit token-wise activation reuseFLUX.1 Kontext · Multi-round IIE benchmarkspeedup: 1.000 -> 1.915Improved
add CacheEdit token-wise activation reuseQwen-Image-Edit · Multi-round IIE benchmarkspeedup: 1.000 -> 1.613Improved
CacheEdit instead of RegionEFLUX.1 Kontext · Multi-round IIE benchmarkPSNR vs vanilla: 29.103 -> 39.112Improved
CacheEdit instead of RegionEQwen-Image-Edit · Multi-round IIE benchmarkLPIPS vs vanilla: 0.098 -> 0.063Improved
add CacheEdit reuse (vs full recomputation)FLUX.1 Kontext · Multi-round IIE benchmarkGPT-4o overall score (G-O): 6.336 -> 6.814Improved
add CacheEdit reuse (vs full recomputation)FLUX.1 Kontext · Multi-round IIE benchmarkGPT-4o semantic consistency (G-SC): 7.566 -> 7.816Improved
add CacheEdit reuse (vs full recomputation)Qwen-Image-Edit · Multi-round IIE benchmarkGPT-4o overall score (G-O): 7.384 -> 7.889Improved
add CacheEdit reuse (vs full recomputation)Qwen-Image-Edit · Multi-round IIE benchmarkGPT-4o semantic consistency (G-SC): 8.077 -> 8.358Improved
raise similarity threshold from 0.80 toward 1.00 (ablation)CacheEdit · FLUX.1 Kontextreconstruction fidelity (SSIM/LPIPS/PSNR): 0.80 -> 1.00Improved
lower similarity threshold (more editable tokens) (ablation)CacheEdit · FLUX.1 KontextspeedupHurt
broader edit type (style transfer vs subject edit) (ablation)CacheEdit · FLUX.1 Kontextspeedup: 2.793 -> 1.156Hurt
broader edit type (background change vs subject edit) (ablation)CacheEdit · Qwen-Image-Editspeedup: 2.376 -> 1.157Hurt

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy