A 4-step distilled version of Iris-3B. Pixel-space text-to-image with guidance baked into the weights: one forward pass per step, no VAE.
