Hanqiu Li Cai

The authors pretrain Iris-3B, a 3B-parameter pixel-space text-to-image transformer trained from scratch through a 256→512→1024 curriculum with v-prediction and DINOv2 REPA, and convert the latent FLUX.2 Klein base 4B to pixel space; after fine-tuning both pixel backbones for monocular depth and image restoration, they find no significant improvement from a pixel-space generative prior: on depth Iris-3B is level with latent FLUX.2 Klein while the converted model trails it, and on DIV2K restoration neither pixel model beats a latent FLUX.2 Klein fine-tune.

The authors pretrain Iris-3B, a 3B-parameter pixel-space text-to-image transformer trained from scratch through a 256→512→1024 curriculum with v-prediction and DINOv2 REPA, and convert the latent FLUX.2 Klein base 4B to pixel space; after fine-tuning both pixel backbones for monocular depth and image restoration, they find no significant improvement from a pixel-space generative prior: on depth Iris-3B is level with latent FLUX.2 Klein while the converted model trails it, and on DIV2K restoration neither pixel model beats a latent FLUX.2 Klein fine-tune.