Research map / Image generation / Text-to-image diffusion
Adversarial Perturbation Shield: Preventing Concept Bleed-through in Continual Learning of Personalized Generative Models
Ziwen Lan, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama
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 adversarial perturbation to training images | C-LoRA · continual personalization | A_mmd: 2.95 -> 1.99 | Improved |
| add adversarial perturbation to training images | STAMINA · continual personalization | A_mmd: 2.34 -> 1.92 | Improved |
| add adversarial perturbation to training images | CIDM · continual personalization | CLIP score: 0.85 -> 0.87 | Improved |
| add adversarial perturbation to training images | CIDM · continual personalization | A_mmd: 2.11 -> 1.98 | Improved |
| add adversarial perturbation to training images | C-LoRA · continual personalization | F_mmd (forgetting): 0.23 -> 0.32 | Hurt |
| add adversarial perturbation to training images | CIDM · continual personalization | F_mmd (forgetting): 0.22 -> 0.22 | No clear effect |
| add adversarial perturbation to training images | STAMINA · continual personalization | F_mmd (forgetting): 0.27 -> 0.26 | No clear effect |
| use a semantically dissimilar target image (dog) instead of a close one (vase) for perturbation (ablation) | CIDM + perturbation · table -> chair sequence | A_mmd: 2.82 -> 1.92 | Improved |
| add perturbation with a bread target image (ablation) | CIDM · table -> chair sequence | A_mmd: 3.12 -> 1.91 | Improved |
| use perturbation instead of selective amnesia | C-LoRA · table -> chair sequence | Table CLIP score (previous concept): 0.094 -> 0.822 | Improved |
| use perturbation instead of selective amnesia | C-LoRA · table -> chair sequence | Chair CLIP score (new concept): 0.851 -> 0.847 | No clear effect |
| train sequence up to ten concepts with perturbation | STAMINA/CIDM · 10-concept sequence | image quality after ten concepts | Improved |
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.
- InstantBooth — InstantBooth: Personalized Text-to-Image Generation without Test-Time Finetuning (CVPR 2024) · better on 3 of 3 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
- How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization? (NeurIPS 2024 · similarity 0.86)
- Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.86)
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR 2024 · similarity 0.85)
- APT: Adaptive Personalized Training for Diffusion Models with Limited Data (CVPR 2025 · similarity 0.85)
- CAT: Contrastive Adversarial Training for Evaluating the Robustness of Protective Perturbations in Latent Diffusion Models (ICML 2025 · similarity 0.85)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.85)
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.
