Search papers, labs, and topics across Lattice.
This study investigates the preservation of training gradient signals when training diffusion transformers on downscaled images, revealing that while downscaling can maintain the surviving signal at high noise levels, it also introduces a noise-independent floor that affects gradient fidelity. The authors identify a specific range of noise levels where downscaled gradients closely approximate native gradients, which is not predicted by existing spectral theories. Implementing this insight, they demonstrate a 14.6% reduction in training time for LoRA adapters without sacrificing performance, showcasing a practical optimization for image generation tasks.
Downscaled training can yield gradients nearly identical to native ones, but only within a specific noise window that defies spectral predictions.
Diffusion transformers deliver strong image generation, but their training cost grows superlinearly with resolution. Recent work justifies training or sampling at reduced resolution on a spectral premise: at high noise, a downscaled latent preserves almost the full surviving signal. Whether a downscaled step also preserves the native training gradient signal, however, has remained unresolved. We reduce how that signal changes under downscaling to two terms: a noise-dependent term governed by the downscale ratio, which decays at high noise as the spectral premise predicts, and a \sigma-independent floor governed by the target grid's absolute token count, carried by the compute graph itself and removed by no noise level. The measured (route, \sigma) map corroborates the account and uncovers structure the spectral picture cannot express: on the 1024->768 route, a window (0.65<\sigma<0.95), predicted by no spectral criterion at any tolerance, where the downscaled gradient stays within a small margin of the native one. Training LoRA adapters with downscaled steps restricted to the routes and noise windows the map validates reduces training time by 14.6% at a fixed step budget while remaining near-native in weight space. Code is available at https://github.com/sorryhyun/anima_lora.
