← 研踪 PaperTrailarXiv:2603.26783 · v8
Can We Change "Stroke Size" to Improve Flow Matching?
首次提交 2026-03-25 14:03
版本更新 2026-11-08 11:07
雷达观测 2026-11-10 12:16
摘要
Deep-learning-based image generation models are useful but difficult to train. Inspired by the varying stroke sizes used in human painting, we hypothesize that applying ``rough strokes'' can attenuate prediction biases and improve flow-matching image generation. We develop MultiStroke, which progressively varies the spatial granularity of the clean-signal component in the velocity regression target. A lightweight corrector aligns adjacent roughness levels during sampling. On ImageNet-1K, MultiStroke reduces FID by $22.5\%$ relative to a matched flow-matching baseline, despite the limited corrector capacity. Code is available at https://github.com/DaChengRuoQue/MultiStroke-Code. A quick video explanation is available at https://dachengruoque.github.io/MultiStroke-Web/.
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v8 · 2026-11-08 11:07 · 最新观测版本
Can We Change "Stroke Size" to Improve Flow Matching?
Deep-learning-based image generation models are useful but difficult to train. Inspired by the varying stroke sizes used in human painting, we hypothesize that applying ``rough strokes'' can attenuate prediction biases and improve flow-matching image generation. We develop MultiStroke, which progressively varies the spatial granularity of the clean-signal component in the velocity regression target. A lightweight corrector aligns adjacent roughness levels during sampling. On ImageNet-1K, MultiStroke reduces FID by $22.5\%$ relative to a matched flow-matching baseline, despite the limited corrector capacity. Code is available at https://github.com/DaChengRuoQue/MultiStroke-Code. A quick video explanation is available at https://dachengruoque.github.io/MultiStroke-Web/.
查看此版本解读 →查看此版本原文 ↗