AI & Creative Tools
Merging 1,080 LoRAs by Preference: HimmBO Cuts the Image Generations by Over 10x
Adapters sharing a base model can be merged with continuous coefficients, which turns a 100K-adapter library into a continuous design space nobody can search by hand. HimmBO screens first on images that already exist, then optimises coefficients in the recovered subspace — reaching a target similarity with under a tenth of the image generations, and holding up at 1,080 adapters where prior methods degrade.