The authors introduce Pristine Attention, which divides attention layers into three independently parameterized branches: safe, unsafe and shared.
Models can isolate restricted image concepts during pretraining
A three-branch attention architecture suppresses designated concepts by disconnecting one branch while preserving useful knowledge learned from restricted images.
Big Tech
Anirudh Sundara Rajan · Krishna Kumar Singh · Richard Zhang · Yong Jae Lee
University of Wisconsin–Madison · Adobe Research
Research Digest··2 min read
Rajan and colleagues redesign text-to-image pretraining so that knowledge associated with designated unsafe concepts is directed into a removable component rather than erased afterward.
Why this paper
From Adobe Research and University of Wisconsin–Madison
In one line
Unsafe concept knowledge in text-to-image models can be isolated during pretraining into a detachable branch, preserving safe-data quality while suppressing restricted generation.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·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.
§