Pyramid-JiT: Low-Res Drafts, High-Res Images
Summary
Pyramid-JiT (P-JiT) is a decoder-only, pixel-space text-to-image model that abandons VAEs and learns by predicting images at multiple resolutions along the DiT trunk. It achieves substantially faster training and inference, reaching performance close to or better than prior models (Linum v2, JiT-DDT) with significantly fewer training samples and GPU hours, aided by multi-readout and multi-patchifications. The work includes architecture details, training regime, image samples, and prompt-following benchmarks, and is released under Apache 2.0.