AnalysisAI ModelsOctober 8, 2026

Iris-3B: 3B-parameter pixel-space text-to-image diffusion model

Read original source →arxiv.org

Iris-3B is a 3B-parameter pixel-space text-to-image transformer pretrained from scratch, avoiding the lossy VAE used by latent diffusion models. The paper tests whether pixel-space backbones hold an advantage on downstream tasks where fine-grained detail matters, covering both training and conversion routes.

1 source

More stories today

Open the live feed