TRAINING A TEXT-TO-IMAGE MODEL FOR $21K

Summary from the TLDR AI newsletter

Linium is a two-person research lab that is trying to train its own generative video models entirely from scratch, aiming to create tools to democratize access to animation. Pyramid-JiT (P-JiT) is a pixel-space architecture that achieves aggressive token reduction through a noise schedule that helps pixel-space models better learn fine-grained detail. This study looks at how models trained on the architecture perform when trained for longer as a pressure test along the way towards larger models and video. Examples of images generated by the resulting model are available.