Understanding LoRA Rank Trade-offs in Diffusion Model Fine-Tuning

论文概要

研究领域: ML 作者: Iman Khazrak, Narges Nejad, Mostafa M. Rezaee 发布时间: 2026-09-11 arXiv: 2509.05825

中文摘要

为扩散模型微调选择LoRA秩需要在质量和计算成本间权衡。本文在CIFAR-10上进行对照研究，使用DDPM U-Net，秩取{2,4,8,16,32}，固定优化设置，并采用可复现的本地文件夹pytorch-fid协议。我们报告FID、可训练参数、运行时间和GPU内存，然后通过扩展预算DDPM运行（20轮；秩4/8/16）和Tiny DiT骨干网络（10轮；秩4/8/16）验证趋势。结果显示中等秩最高效：秩4取得最佳DDPM FID (124.1380)，秩8接近(124.2136)，更高秩尽管适配成本更大但收益有限。这些发现支持在固定训练预算下，中小秩作为实用默认值。

原文摘要

Selecting LoRA rank for diffusion fine-tuning requires balancing quality and compute cost. We present a controlled study on CIFAR-10 using a DDPM U-Net with ranks {2,4,8,16,32}, fixed optimization settings, and a reproducible local-folder pytorch-fid protocol. We report FID, trainable parameters, runtime, and GPU memory, then validate trends with extended-budget DDPM runs (20 epochs; ranks 4/8/16) and a Tiny DiT backbone (10 epochs; ranks 4/8/16). Results show moderate ranks are most efficient: rank 4 achieves the best DDPM FID (124.1380), rank 8 is close (124.2136), and higher ranks provide limited gains despite larger adaptation cost. These findings support small-to-moderate ranks as practical defaults under fixed training budgets.

#论文 #arXiv #ML #小凯