Iris-3B: 3B pixel-space diffusion model released, with negative transfer result
Researchers at SperidLabs released Iris-3B, a 3B-parameter pixel-space text-to-image transformer pretrained from scratch through a 256→512→1024 curriculum, together with its weights and training code; they also converted the pretrained latent model FLUX.2 Klein base 4B into pixel space. Fine-tuning both families for monocular depth estimation and image restoration/super-resolution, they found no significant benefit from a pixel-space generative prior: Iris-3B matched the latent FLUX.2 Klein on depth while the converted version trailed it, and on 4× DIV2K restoration neither pixel model beat the latent fine-tune. Still, the authors report that at 1024² Iris-3B's text-to-image quality matches Qwen-Image under the official evaluators, showing pixel-space pretraining scales to 3B parameters.
The negative result undercuts claims that pixel-space diffusion is inherently better for fine-grained downstream tasks, while showing the approach can scale to 3B parameters with competitive text-to-image quality.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.09450 (cs)
[Submitted on 7 Oct 2026]
Title:Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning
Authors:Hanqiu Li Cai (SperidLabs), Chema Garabito (SperidLabs)
View a PDF of the paper titled Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning, by Hanqiu Li Cai (SperidLabs) and 1 other authors