Do adaptive optimizers still work? Reddit users doubt it for flow-matching models

Inevitable_Pen9043 · reddit · 2026-11-22

A Reddit post questions whether adaptive optimizers suit the new generation of image models. In the SDXL era they were crucial, but for flow-matching models (most releases after Flux), adaptive optimizers seem to underestimate learning rates, and fixed learning rates often work better — their curves no longer follow diffusion-model logic.

Original post →

More from Multimodal

Multimodal channel →