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 ($\mathbf{v}$) with clean image prediction (${\mathbf{x}}_0$). In this work, we revisit this choice and introduce **Mꜰʟᴏᴡ**, an adaptive parameterization that learns to interpolate between ${\mathbf{x}}_0$ and $\mathbf{v}$ prediction across principal directions of the data, which can be learned end-to-end with diffusion model training. We find that the preferred parameterization varies systematically with the data representation and model capacity. Notably, increasing model capacity shifts preference toward $\mathbf{v}$ prediction from $x_0$ prediction. We evaluate Mꜰʟᴏᴡ across ImageNet, UCF-101 video, and large-scale text-to-image generation, demonstrating competitive or improved generative performance across diverse settings. In addition, we find that vanilla pixel diffusion transformers, irrespective of parameterization, retain a persistent high-frequency deficit relative to real images. To address this, we introduce **_ConvRefiner_**, a lightweight convolutional module trained post-hoc with adversarial loss while keeping the diffusion backbone frozen. With only a small amount of additional training, _ConvRefiner_ restores the high-frequency statistics of generated images while preserving FID.