Speridlabs · Pixel-Space Diffusion

Iris-3B

A 3B-parameter text-to-image model, released 7 October, that generates pixels directly with no VAE or latent space, at around one megapixel. The architecture uses 8 dual-stream and 16 single-stream blocks over 16x16 pixel patches. At 1024x1024 it scores 0.798 on GenEval, 86.52 on DPG, 0.857 on LongText and 0.540 on OneIG, vendor-reported, and the authors claim it is competitive with latent models at a fraction of the training compute. Fine-tuned without architecture changes, the same model also does depth estimation and image restoration and upscaling. Weights, code and a demo are public under Apache 2.0; paper arXiv 2610.09450.