Training a Text-to-Image Model for $21K
Linum released code and weights for a 2B-parameter pixel-space text-to-image model under Apache 2.0, reporting $21,038 in compute for the training and post-training runs. That figure uses an assumed $2.50 per H100-hour and does not establish the full cost of the research, data or development.
The run used 660 million samples and a separate 4B text encoder. The authors show failures with full bodies, multiple entities and interactions, and say it cannot reliably render text because text-heavy images were filtered from training. HN readers asked about the training data and how representative the showcase images are.