Two new research papers introduce methods for concept erasure in autoregressive image generation models, addressing concerns about misuse and the creation of unsafe content. The first paper, Obliviate, proposes a guidance-based approach for these models, focusing on visual token distributions and trajectory-level updates. It demonstrates significant reductions in nudity on a benchmark dataset while preserving overall model utility. The second paper, ScaleErasure, offers an inference-time intervention method that targets specific logits to overcome semantic entanglement in next-scale autoregressive models, showing improved precision in concept erasure while maintaining generative capabilities. AI
IMPACT These methods could enhance safety and control in generative AI, reducing the potential for misuse in autoregressive image generation.
RANK_REASON Two academic papers published on arXiv introducing novel methods for concept erasure in image generation models.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
