CollectionLoRA: Collecting 50 Effects in 1 LoRA via Multi-Teacher On-Policy Distillation
Researchers have developed CollectionLoRA, a new framework that distills up to 50 distinct visual effects from individual Low-Rank Adaptation (LoRA) models into a single LoRA. This approach aims to reduce deployment overhead and prevent concept bleeding and style degradation that occur when multiple LoRAs are cascaded. The method utilizes a probabilistic routing mechanism, asymmetric prompting, and a coarse-to-fine distillation objective to isolate concepts and maintain fidelity. AI
IMPACT Consolidates multiple visual effect LoRAs into one, potentially reducing inference costs and simplifying deployment for customized image editing.