safety security
Fewer Image Steps Made Unsafe Concepts Harder to Remove
Across tested distilled text-to-image systems, lower step counts raised unsafe-generation rates; DICE reduced FLUX Schnell's measured rate to 0.046.
Summary
Across tested distilled text-to-image systems, lower step counts raised unsafe-generation rates; DICE reduced FLUX Schnell's measured rate to 0.046.
Fast text-to-image models compress generation into only a few denoising steps, but the safety behavior of that compression has received less scrutiny. This study reports that several distilled few-step models generate unsafe concepts more often than their teacher models and become harder to modify with existing concept-erasure methods as step counts fall. The authors attribute part of the failure to fragile generation trajectories and the way guidance is distilled, then introduce DICE, which aligns erasure with the originating distillation method. On their FLUX Schnell evaluation, DICE reduced the unsafe-generation rate to 0.046, compared with at least 0.161 for tested prior erasure methods, while largely preserving benign quality. Those rates depend on the study's prompts, detectors and models.
Why it matters
Across tested distilled text-to-image systems, lower step counts raised unsafe-generation rates; DICE reduced FLUX Schnell's measured rate to 0.046.
Limits and context
- Fast text-to-image models compress generation into only a few denoising steps, but the safety behavior of that compression has received less scrutiny.
Key claims
Across tested distilled text-to-image systems, lower step counts raised unsafe-generation rates; DICE reduced FLUX Schnell's measured rate to 0.046.
Qualification: Fast text-to-image models compress generation into only a few denoising steps, but the safety behavior of that compression has received less scrutiny.
Evidence: source-2026-10-28-011
Sources
- arXiv preprint 2610.36665, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.