New TEA framework enhances concept erasure in text-to-image models

Researchers have developed a new framework called TEA (Text Encoder Alignment) to improve concept erasure in text-to-image diffusion models. This method fine-tunes only the text encoder, keeping the generative backbone frozen, to make concept-containing prompts indistinguishable from safe ones. TEA offers state-of-the-art robustness against adversarial attacks on models like Stable Diffusion v1.4 and v3.5, while also preserving the quality of benign generations and introducing no inference-time overhead. AI

IMPACT Improves safety and robustness of text-to-image models against misuse.

RANK_REASON This is a research paper detailing a new method for concept erasure in text-to-image models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Alireza Dehghanpour Farashah
• arXiv
• DagsHub
• Hugging Face
• Rectified Flow transformer
• Stable Diffusion v1.4
• Stable Diffusion v3.5
• T5 Text To Text Transfer Transformer
• TEA
• Text Encoder Alignment

• paper
• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →