PixelU: A U-Shaped Transformer for Efficient End-to-End Pixel Diffusion

A new minimalist transformer architecture called PixelU achieves state-of-the-art image generation quality on the ImageNet benchmark while using roughly one-third the computation of a leading competitor, according to a preprint posted to arXiv on June 26, 2026 [1]. The model, described in a paper submitted to the Computer Vision and Pattern Recognition section of the repository, takes an end-to-end approach to pixel-space diffusion, a technique that works directly with raw image pixels rather than the compressed representations used by Latent Diffusion Models [1]. The authors argue that existing pixel-space systems have relied on complex auxiliary decoders primarily to compensate for optimization difficulties tied to velocity prediction. Under a clean-data paradigm using x-prediction, the researchers contend, those decoders become redundant [1]. PixelU replaces auxiliary decoders with zero-cost skip connections that function as an "information highway," routing uncorrupted high-frequency spatial details from shallow layers directly to deeper layers of the network [1]. To allow the backbone to concentrate on low-frequency semantics, the design incorporates a constant-channel spatial down-sampling mechanism that acts as a natural low-pass filter, compressing deep features into a compact, low-frequency semantic manifold [1]. The architecture belongs to the broader family of vision transformers, which decompose an input image into a series of patches, serialize each patch into a vector, and process the resulting embeddings with a transformer encoder [3]. Vision transformers were developed as alternatives to convolutional neural networks and have been shown to possess higher capacity, though they are generally less data-efficient [3]. On the ImageNet dataset at 256×256 resolution, PixelU recorded a Fréchet Inception Distance (FID) of 1.63; at 512×512, the score was 1.92 [1]. The authors report that these results surpass recent pixel-space methods and outperform the strong baseline JiT-G while using only about one-third of its computational cost [1]. The paper appeared on arXiv, an open-access repository that hosts electronic preprints across mathematics, physics, computer science, and related fields [10]. As of November 2024, the repository was receiving approximately 24,000 submissions per month [10]. The preprint has not yet undergone peer review [10].