RFD 1042: Model image for krea2_turbo_text_to_image

Problem

Krea 2 Turbo is the largest model in the catalog that the project runs. It needs 33.8 GB in bf16, which is 29 percent of the whole catalog.

It is also four models in one folder: a backbone, two text encoders, and a VAE. A model image that loads all four at once wastes memory, because the text encoders finish before the backbone starts.

Decision

Package it as one model image, and stage the loads. The text encoders load, they run, and they unload. The backbone then loads.

Quantize to Q4_K_M. RFD 1027 selects that format for both text-to-image models, and it drops this one from 33.8 GB to 9.30 GB.

See DETAILS.md for the model’s per-part memory, the predict() interface, and the disk-size trap in the weight folder.

Details

The measurements and the retractions