Rethinking the design space of Vanilla Pixel Diffusion Transformers
Advances diffusion model architecture with adaptive parameterization and high-frequency refinement for researchers working on image generation.
AI Summary
Researchers introduce MFlow, an adaptive parameterization for pixel diffusion transformers that learns to interpolate between x₀- and u-prediction, plus ConvRefiner to address high-frequency deficits.
Excerpt
Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and separately trained autoencoder. JiT shows that a vanilla diffusion transformer can generate directly in pixel space by replacing the commonly used velocity prediction ($\boldsymbol{u}$-prediction) with clean image prediction ($\boldsymbol{x_0}$-prediction). In this work, we revisit this choice and introduce MFlow, an adaptive parameterization that learns