Text-to-image diffusion models are often distilled into few-step variants for fast inference, but these models can still generate harmful content.
Consistency error enables direct unlearning in few-step distilled diffusion models
The proposed CePU framework replaces noise-prediction rewards with consistency error, achieving effective concept removal in distilled models without re-distillation.
Big Tech
Gaurav Patel · Jun Fang · Greg Ver Steeg · Qiang Qiu · Sravan Sripada
Amazon AGI · Purdue University
Research Digest··2 min read
The authors introduce CePU, a preference-driven unlearning method tailored to few-step distilled (FSD) text-to-image diffusion models.
Why this paper
From Amazon AGI and Purdue University
In one line
Using consistency error instead of noise prediction error achieves effective unlearning in few step distilled text to image models.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§