The authors add an adversarial discriminator loss to pretrained pixel (DeCo, PixelGen) at low-noise timesteps, keeping the original diffusion objective. This post-training step consistently improves distribution fidelity, coverage, prompt alignment, and perceptual quality. For example, on DeCo, FID drops from 33.27 to 28.59 and recall rises from 0.361 to 0.406.
The improvement comes from correcting a systematic underproduction of high-frequency content. The GAN restores missing spectral power without the domain shift seen with perceptual losses. In contrast, the same procedure fails on latent diffusion models because the frozen decoder attenuates corrections to final pixels.
This work identifies a simple, architecture-preserving fix for a known pixel diffusion limitation and explains why the same trick does not work in latent diffusion, highlighting direct output access as a key factor.