Adversarial Training for Pixel Diffusion
Adversarial post-training restores missing high-frequency image detail in pixel diffusion models without changing sampling.
The paper adds an adversarial loss on top of the original diffusion or flow-matching objective, applied away from high-noise timesteps. Across two pixel backbones, the authors report gains in fidelity, coverage, prompt alignment, and perceptual quality. Their analysis says the baseline models underproduce natural-image high-frequency content, and the adversarial step restores that spectral power. Similar tests on the latent diffusion setups they tried did not produce comparable gains. HF Daily Papers' note
The paper adds an adversarial loss on top of the original diffusion or flow-matching objective, applied away from high-noise timesteps. Across two pixel backbones, the authors report gains in fidelity, coverage, prompt alignment, and perceptual quality. Their analysis says the baseline models underproduce natural-image high-frequency content, and the adversarial step restores that spectral power. Similar tests on the latent diffusion setups they tried did not produce comparable gains. HF Daily Papers' note
score 4