Training Text-to-Image Models 3.6× Faster

Linum.ai has developed a method to train text-to-image models 3.6x faster by optimizing token reduction and prediction. They use 32x32 patches for 512x512 images, resulting in 256 tokens, a technique similar to vision transformers (ViT). Linum v2 employs a VAE for 8x8 compression and 16-dimensional latents, further applying 2x2 patchification for 16x16 token compression and 64-dimensional latents. By switching to x-prediction instead of v-prediction, the model can dedicate its full capacity to the low-dimensional signal, improving efficiency.

Unlike previous methods that focused on noise prediction, Linum.ai's approach switches to x-prediction, allowing the model to dedicate its full capacity to the low-dimensional signal.