Training Text-to-Image Models 3.6× Faster
Summary
Field Notes from Linum detail Training Text-to-Image Models 3.6× Faster, introducing JiT-DDT, a pixel-space encoder–decoder approach that eliminates VAE bottlenecks by unifying compression and generation in a single model. The post covers hitting the VAE compression wall, aggressive token compression, extending JiT for text-to-image, recovering fine details with PixelREPA, architecture refinements, and a From Linum v2 to JiT-DDT progression, supported by extensive comparisons and ablations. It presents practical benchmarks (3.6× faster training with 4× the pixels) and a rich appendix of image prompts comparing Linum v2 and JiT-DDT.