AI Paper Atlas

Research map / Image generation / Text-to-image diffusion

Co-occurring Associated REtained concepts in Diffusion Unlearning

Miso Kim, Georu Lee, Yunji Kim, Hoki Kim, Jinseong Park, Woojin Lee

ICLR 2026 · Image generation › Text-to-image diffusion · approach: Concept erasure in diffusion models

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CARE-set retain/erase losses (ReCARE) to STEREO-style unlearningSTEREO · Nudity unlearningCARE score: 0.11 -> 0.94Improved
add CARE-set retain/erase losses (ReCARE) to STEREO-style unlearningSTEREO · Nudity unlearning (CCE attack)ASR (%): 19.55 -> 11.14Improved
add global clustering to CARE-set refinement (ablation)ReCARE · Nudity unlearning (CCE attack)ASR (%): 25.00 -> 11.14Improved
add intra-cluster refinement (ablation)ReCARE · Nudity unlearning (CCE attack)ASR (%): 16.36 -> 11.14Improved
add both refinement stages to CARE-set (ablation)ReCARE · Nudity unlearningCARE score: 0.88 -> 0.94Improved
vary number of k-means clusters n from 4 to 6 (ablation)ReCARE · Nudity unlearning (CCE attack)ASR (%): 15.36 -> 11.14No clear effect
vary pruning strictness alpha (0.005 to 0.015) (ablation)ReCARE · Nudity unlearningCARE score: 0.94 -> 0.90No clear effect
vary Top-K tokens per image (30 to 70) (ablation)ReCARE · Nudity unlearningCARE score: 0.97 / 0.94 / 0.96No clear effect
increase retain weight lambda from 1.0 to 1.5 (ablation)ReCARE · Nudity unlearning (CCE attack)ASR (%): 11.14 -> 30.91Hurt
decrease retain weight lambda from 1.0 to 0.5 (ablation)ReCARE · Nudity unlearningFID: 13.85 -> 15.50Hurt
replace CLIP with SigLIP in CARE score evaluationCARE score · Nudity unlearningmethod rankingNo clear effect
use manual anchor 'figure' instead of 'person' for STEREOSTEREO · Nudity unlearninghuman detection rate: 0.64 -> 0.92Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy