Iris-3B Preprint Explores Pixel-Space Diffusion Scaling and Downstream Tasks
The October 2026 arXiv preprint presents Iris-3B, a 3B-parameter pixel-space text-to-image transformer pretrained from scratch, and reports negative results when testing…
- Iris-3B matches latent FLUX.2 Klein on monocular depth after matched fine-tuning but shows no significant advantage on 4x DIV2K restoration.
- The preprint releases Iris-3B weights and training code to support further research on pixel-space generation at 1024² resolution.
- Pixel-space pretraining with the PiT head scales to 3B parameters and achieves text-to-image quality matching Qwen-Image on OneIG at 1024².
What the preprint investigates
The abstract-only evidence describes testing whether pixel-space diffusion models deliver advantages on downstream tasks that require fine-grained detail by avoiding the lossy VAE used in latent models.
Authors pretrained Iris-3B from scratch through a 256-to-512-to-1024 resolution curriculum after performing ablations at 256² on prediction target and representation alignment. They also converted the pretrained latent FLUX.2 Klein 4B model to pixel space.
Key experimental outcomes
According to the abstract, fine-tuning both pixel-space approaches and the original latent model for monocular depth estimation with one matched direct-regression recipe showed no significant improvement from a pixel-space generative prior.
Iris-3B performed at the same level as latent FLUX.2 Klein while the converted pixel FLUX.2 Klein fell behind. On 4× DIV2K restoration neither pixel model outperformed a latent FLUX.2 Klein fine-tune, with the converted version trailing slightly.
The authors document failure modes and remaining confounds that may explain this negative result.
Base model performance and release
Despite the downstream findings, Iris-3B demonstrates that pixel-space pretraining using the pixel-transformer head of PixelDiT scales successfully to 3B parameters.
It reaches text-to-image quality competitive with latent models, matching Qwen-Image on the OneIG benchmark under official evaluators at 1024² resolution.
The preprint releases Iris-3B weights and training code, providing a reproducible foundation for continued work on pixel-space generation.
