Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning

Iris-3B is a 3B-parameter pixel-space text-to-image transformer pretrained using a 256 to 1024 resolution curriculum. While designed to avoid VAE compression losses, empirical testing shows the model provides no significant improvement over latent-space generative priors for monocular depth estimation or image restoration tasks.