Blue Mage is an open research initiative focused on the learning dynamics of text-to-image models. We aim to make training trajectories available as research resources, so that researchers can examine how model behavior develops across training time, model size, and data scale.
Our upcoming release covers MMDiT models with approximately 0.5B, 2B, and 3.9B parameters, trained from scratch using GPIC datasets at scales of 1M, 10M, and 100M examples. The collection is being prepared for release alongside an arXiv paper.
The release is designed to include intermediate model checkpoints, EMA weights, optimizer states, data-order information, and supporting training artifacts such as latents and materialized noise. Availability and the contents of each release will be documented in the corresponding repositories. The training pipeline supports bitwise-resumable continuation. Together with the intermediate checkpoints and training records, this enables researchers to restore runs and investigate their subsequent trajectories.
We are interested in how generative models acquire capabilities, generalize, memorize, and respond to changes in model and data scale. By opening the training process to inspection, we hope to support controlled comparisons and new analyses that are difficult to conduct using final model weights alone.
Hugging Face: Here!
ModelScope: https://www.modelscope.ai/organization/BlueMage