Rethinking the design space of Vanilla Pixel Diffusion Transformers
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 ($\boldsymb
Reading in Plain English mode — toggle below to switch
Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and separately trained autoencoder. JiT shows that a vanilla diffusion transformer can generate directly in pixel space by replacing the commonly used velocity prediction ($\boldsymbol{u}$-prediction) with clean image prediction ($\boldsymb
This paper explores improvements in pixel-space diffusion models, which are becoming a strong alternative to latent diffusion. Understanding adaptive parameterization like MFlow is crucial for ML engineers working on high-quality image generation and efficient model design, as it directly impacts model performance and computational cost.