AI Paper Atlas

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

Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models

Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, Yu-Gang Jiang

ECCV 2024 · Image generation › Text-to-image diffusion · approach: Concept erasure in diffusion models · 64 citations (OpenAlex)

Read the paper

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
derive and erase new embeddings iteratively instead of single-pass UCE-style editUCE · I2P nuditytotal nude parts detected: 186 -> 66Improved
apply RECE instead of UCE against Ring-A-BellUCE · Ring-A-Bellattack success rate (%): 33.10 -> 13.38Improved
RECE versus CA under white-box P4D attackRECE · P4Dattack success rate (%): 60.56 -> 64.79Hurt
RECE vs UCE on COCO-30k image qualityUCE · COCO-30kFID: 1.82 -> 2.82Hurt
run derive-erase for more epochs (epochs 3-4)RECE · nudity erasuremodel generation abilityHurt
increase regularization coefficient lambdaRECE · COCO-30k validation subsetCLIP scoreImproved
increase regularization coefficient lambdaRECE · I2Pnude parts detectedNo clear effect
apply RECE for male nudity erasureRECE · I2Pfemale/male nudity ratio: female 204 -> 8, male 56 -> 10Mixed
apply RECE for artist style removal compared to UCEUCE · Erase Van GoghLPIPS_d: 0.20 -> 0.23Improved
apply RECE for object removal compared to UCEUCE · Imagenette object removalaccuracy of erased class (%): 2.6 -> 0.3Improved

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