A new preprint on arXiv, listed as arXiv:2610.23854, re-examines the design space of pixel-space diffusion transformers. The authors argue that these models have recently re-emerged as a compelling alternative to latent diffusion, which typically compresses data into a latent space before generating samples.

The key advantage, according to the abstract, is that pixel-space models avoid both the lossy latent bottleneck and the need for a separately trained autoencoder. The work, referred to as JiT, proposes that a vanilla diffusion transformer can operate directly in pixel space.

Because the source abstract is cut off mid-sentence, the specific results or benchmarks are not available. This digest is based solely on the provided excerpt; no further claims about performance can be verified from the source.