Few-step image models raise unsafe output; DICE cuts it
Original title: Haste Makes Waste? Towards Safe Few-Step Text-to-Image Models
On FLUX Schnell, DICE lowers the unsafe-generation rate to 0.046, versus at least 0.161 for prior erasure methods tested.
A paper finds few-step distilled text-to-image models are often less safe than their teachers, and that fewer inference steps raise unsafe generation rates. Existing concept-erasure methods are markedly less effective after acceleration, which the authors link to fragile few-step generation and how guidance is distilled. They propose DICE, a post-distillation erasure method aligned with the distillation procedure. On FLUX Schnell, DICE reduces the unsafe-generation rate to 0.046 compared with at least 0.161 for prior methods they test, while mostly preserving benign image quality.
- Fewer inference steps consistently raised unsafe generation rates.
- Concept erasure was harder after acceleration than on teacher models.
- DICE matches latent construction and preservation loss to the distillation method.
- Results are reported for FLUX Schnell in the authors' evaluation.
- Products
- DICE
- AI models
- FLUX Schnell
From the original article
1 min read at arxiv.orgWhile few-step distilled text-to-image models have gained traction due to their fast generation times, the speed-up implications on their unsafe generation rates remain unknown. We uncover an inherent tension between speed and safety: few-step models are often more unsafe than their teachers, and reducing the number of inference steps consistently raises unsafe generation rates.
By Lior Ziv · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.