Adobe Research Shows Adversarial Post-Training Restores Missing High-Frequency Detail in Pixel Diffusion

adobe-research · hf · 2026-09-30

Adobe Research presents the first systematic study of adversarial post-training for pixel diffusion: adding an adversarial loss on predicted outputs at non-high-noise timesteps, with architecture and sampling unchanged. Across two pixel backbones it jointly improves fidelity, coverage, prompt alignment, and perceptual quality. Frequency and power-law analyses show original models systematically underproduce high-frequency content, which adversarial training restores—while the same procedure fails on latent diffusion, identifying direct output access to image statistics as the key factor.

Original post →

More from Multimodal

Multimodal channel →