arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.26783cs.CVcs.AI
收藏

我们能否通过改变笔触大小来使扩散更简单?

Can We Change "Stroke Size" to Improve Flow Matching?

National University of Singapore(新加坡国立大学);Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州))

Yunwei Bai, Ying Kiat Tan, Tsuhan Chen

AI总结:

本文研究通过控制笔触大小改变监督目标的粗糙度,以缓解低信噪比下的预测挑战。

摘要:

扩散模型在低信噪比条件下面临挑战,需要在高噪声下进行像素级预测。几何直觉类似于整个油画过程中使用最细的笔触,可能效果不佳。因此,我们研究笔触大小控制作为一种受控干预,改变监督目标、预测和扰动在时间步长上的有效粗糙度,以减轻低信噪比挑战。

英文摘要:

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/.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑