Choosing a LoRA rank for diffusion model fine‑tuning involves a trade‑off between sample quality and computational expense. We conduct a controlled study on CIFAR‑10 using a DDPM U‑Net, evaluating ranks {2,4,8,16,32} while keeping all optimizer settings fixed. Evaluation follows a reproducible local‑folder pytorch-fid protocol, and we report FID, number of trainable parameters, runtime, and GPU memory usage.
To confirm the observed trends, we extend the budget: DDPM runs for 20 epochs with ranks 4/8/16, and a Tiny DiT backbone trained for 10 epochs with the same three ranks. The results indicate that moderate ranks are the most efficient: rank 4 achieves the best DDPM FID (124.1380), rank 8 is very close (124.2136), and higher ranks incur larger adaptation costs with only marginal gains.
Thus, under a fixed training budget, small‑to‑moderate LoRA ranks (e.g., 4 or 8) serve as practical defaults.
Review: The paper supplies systematic empirical evidence that guides practitioners to select LoRA ranks wisely in resource‑constrained settings, avoiding unnecessary overhead from overly large ranks.