Discovering Undisclosed Concepts in LoRA Adapters for Diffusion-Based AIGC Models

Key points

• We therefore formulate concept discovery as searching for the conditioning input that jointly maximizes output coherence, LoRA-induced distinctiveness, and subspace alignment.1
• Its alignment with the learned LoRA subspace provides a lightweight signal for localizing promising candidates before visual verification.1
• Based on this insight, we develop a two-stage framework for text-to-image LoRAs.1
• Subspace-Guided Evolutionary Search (SGES) uses render-free subspace alignment to explore candidate prompts and retain a compact shortlist, while Paired Visual Identification (PVI) renders only these candidates and compares the LoRA and base-model outputs to identify those exhibiting coherent and distinctive generation behavior.1

Sources (1)

• [1]Discovering Undisclosed Concepts in LoRA Adapters for Diffusion-Based AIGC Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 06:00 PM

  Low-Rank Adaptation (LoRA) has made diffusion model personalization affordable and widely shareable, but it also creates moderation risks when harmful or problematic content is not faithfully disclosed.

  We therefore formulate concept discovery as searching for the conditioning input that jointly maximizes output coherence, LoRA-induced distinctiveness, and subspace alignment.

Extractive summary: sentences quoted from the sources.