MFlow: Rethinking the Design Space of Vanilla Pixel Diffusion Transformers

Paper Overview

Research Area: Computer Vision Authors: Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava, Aniruddha Mahapatra Published: 2026-10-16 arXiv: 2610.23854

Summary

Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and the need for a separately trained autoencoder. JiT demonstrated that a vanilla diffusion Transformer can generate directly in pixel space by replacing the commonly used velocity prediction with clean image prediction (\(\boldsymbol{x_0}\)-prediction).

This paper revisits that design choice and proposes MFlow, an adaptive parameterization that learns to interpolate between \(\boldsymbol{x_0}\)-prediction and velocity prediction across the principal directions of the data, trained end-to-end jointly with diffusion training. Key findings include:

• The preferred parameterization varies systematically with data representation and model capacity — notably, increasing model capacity shifts the preference toward velocity prediction.
• MFlow is evaluated on ImageNet, UCF-101 video, and large-scale text-to-image generation, achieving competitive or improved generative performance across settings.
• Independently of parameterization, pixel diffusion Transformers exhibit a persistent high-frequency deficit relative to real images.

To address this, the authors introduce ConvRefiner, a lightweight convolutional module trained post hoc with an adversarial loss while freezing the diffusion backbone. With only a small amount of additional training, ConvRefiner restores the high-frequency statistics of generated images while preserving FID.

Abstract (original)

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 to interpolate between \(\boldsymbol{x_0}\)- and \(\boldsymbol{u}\)-prediction across principal directions of the data, learned end-to-end with diffusion training. We find that the preferred parameterization varies systematically with the data representation and model capacity -- notably, increasing mo...

--- *Auto-collected on 2026-10-20*