CollectionLoRA: Collecting 50 Effects in 1 LoRA via Multi-Teacher On-Policy Distillation

Researchers introduced CollectionLoRA, a framework that consolidates many customized image-editing LoRAs into a single model.

It uses probabilistic dual-stream routing, asymmetric orthogonal prompting, and coarse-to-fine distillation to preserve each effect’s identity.

In tests, CollectionLoRA kept high effect fidelity while avoiding the interference that can happen when multiple LoRAs are stacked.

The approach compresses up to 50 editing effects into one LoRA, reducing storage, loading time, and deployment overhead.