Linum's JiT-DDT Trains Text-to-Image Models 3.6× Faster with 4× More Pixels
Training Text-to-Image Models 3.6× Faster
Linum v2 was bottlenecked by its attention context window: a 720p, 5-second clip cost 110K tokens. By replacing the VAE with a unified pixel-space model and switching to x-prediction, Linum's new JiT-DDT architecture trains a text-to-image model with 3.6× fewer GPU-hours while generating images with 4× the pixels. The code and weights are released under Apache 2.0.
By switching to x-prediction, the model can spend its full capacity on the low-dimensional signal, even when D becomes large.
- schopra909
Author here, feel free to drop questions below. Will try to answer to best of my ability!