Research map / Image generation / Text-to-image diffusion
Preference-Calibrated Optimization with Score-Level Distribution Alignment for Text-to-Image Diffusion Model Unlearning
Xiuyuan Wang, Weiming Liu, Hongyu Cai, Xin Gao, Fan Wang, Chaochao Chen, Xiaolin Zheng
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add score-level distribution alignment (beta from 0 to 3e-6) (ablation) | POSDA · UnlearnCanvas Abstractionism | UA/IRA/CRA: beta 0 -> 3e-6 | Improved |
| increase alignment weight beyond optimum (3e-6 to 3e-5) (ablation) | POSDA · UnlearnCanvas Abstractionism | UA: 3e-6 -> 3e-5 | Hurt |
| add dynamic anchors (k = 3) to the naive reward (ablation) | POSDA · UnlearnCanvas Abstractionism | UA convergence and utility | Improved |
| use a single anchor (k = 1) instead of k = 3 (ablation) | POSDA · UnlearnCanvas Abstractionism | UA: k=1 vs k=3 | Hurt |
| use multiple anchors (k = 3) for faster erasure (ablation) | POSDA · UnlearnCanvas Abstractionism | training epochs to converge | Improved |
| POSDA vs best baseline SAeUron on style Van Gogh | POSDA · UnlearnCanvas Van Gogh | Avg (%): 94.89 -> 97.77 | Improved |
| POSDA vs best baseline on object Architectures | POSDA · UnlearnCanvas Architectures | Avg (%): 88.39 -> 98.89 | Improved |
| POSDA vs EDiff on NSFW nudity | POSDA · I2P NER-0.5 | NER (%): 5.1 -> 5.0 | Improved |
| POSDA vs original SD on exposed-part detections | POSDA · I2P total detections | count: 443 -> 63 | Improved |
| POSDA vs original SD on CLIP score | POSDA · MS-COCO | CLIP score: 26.54 -> 26.73 | Improved |
| POSDA vs original SD on FID | POSDA · MS-COCO | FID: 52.00 -> 52.76 | No clear effect |
| ESD erasure vs POSDA on Abstractionism | ESD · UnlearnCanvas Abstractionism | UA (%): 100.00 vs 98.33; Avg 91.43 vs 99.21 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SAeUron — SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders (ICML 2025) · better on 6 of 6 shared results
- SalUn — SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (ICLR 2024) · better on 1 of 1 shared result
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
- Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models (NeurIPS 2026 · similarity 0.94)
- Boosting Alignment for Post-Unlearning Text-to-Image Generative Models (NeurIPS 2024 · similarity 0.91)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.91)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.90)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.90)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.90)
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.
