media creative tools
Adversarial Training Put Fine Texture Back Into Pixel Diffusion
A post-training loss restored missing high-frequency power without changing the sampler or architecture.
Summary
A post-training loss restored missing high-frequency power without changing the sampler or architecture.
The researchers kept a pretrained pixel diffusion model's original objective and added an adversarial loss to predicted outputs at non-high-noise timesteps. Across two pixel backbones, they report joint gains in distribution fidelity, coverage, prompt alignment and perceptual quality, with spectral analysis showing restored high-frequency content. Matched controls were used to test memorization, mode dropping and extra optimization. Under the latent-diffusion configurations they tested, the same procedure did not produce comparable joint gains, marking a reported boundary rather than a general GAN recipe.
Why it matters
A post-training loss restored missing high-frequency power without changing the sampler or architecture.
Limits and context
- Under the latent-diffusion configurations they tested, the same procedure did not produce comparable joint gains, marking a reported boundary rather than a general GAN recipe.
Key claims
A post-training loss restored missing high-frequency power without changing the sampler or architecture.
Qualification: Under the latent-diffusion configurations they tested, the same procedure did not produce comparable joint gains, marking a reported boundary rather than a general GAN recipe.
Evidence: source-2026-09-30-005
Sources
- arXiv preprint 2609.38170arXiv · primary research
Corrections
No corrections have been recorded for this story.