10Eros-Max: Addressing MiniMax-H3 Training Issues via LTX 2.3 Data Migration

To address training issues in the existing MiniMax-H3 model, data extracted from LTX 2.3, Wan 2.2, and Krea 2 was migrated into low-level attention layers. This process involved combining the data in a way that does not compromise H3's visual and audio output quality, aiming to overcome the limitations of the original model.

Specifically, when using TURBO mode, applying the er_sde/simple 6-step or custom 7-step sigma scheduler can effectively remove motion noise. The provided sigma string is [1.00, 0.94, 0.83, 0.72, 0.55, 0.30, 0.10, 0.00], and using these values instead of the default scheduler yields more stable results.

The developer documented the migration code and methodology with the help of Claude, releasing all materials except the scripts themselves as open source. Providing this documentation to an AI agent ensures reproducibility, allowing the model to be migrated in the same manner or its principles to be explained.

The standard H3 community license applies, and depending on the migrated features, the community licenses of LTX 2.3, Wan 2.2, and Krea 2 also apply to those respective parts. This is currently an experimental project dependent on Sulphur H3 tuning, and it is scheduled to be updated as H3 training improves.