Research map / Image generation / Text-to-image diffusion
IREU: Identity-Related Encoder-Only Unlearning for Customized Portrait Generation
Chaoyi Shi, Shanshan Zhang, Jian Yang
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 |
|---|---|---|---|
| perturb only identity-related dimensions instead of global features (IREU vs baseline) | encoder unlearning · CelebA-HQ fidelity, oneID | Delta FID: 4.33 -> 1.02 | Improved |
| perturb only identity-related dimensions instead of global features (IREU vs baseline) | encoder unlearning · CelebA-HQ fidelity, fiveID | Delta FID: 28.37 -> 1.59 | Improved |
| perturb only identity-related dimensions | encoder unlearning · fiveID | ID_other^o: 0.01 -> 0.53 | Improved |
| perturb only identity-related dimensions | encoder unlearning · CelebA-HQ, FastComposer, oneID | Delta ID: 0.48 -> 0.64 | Improved |
| perturb only identity-related dimensions | encoder unlearning · CelebA-HQ, FastComposer, fiveID | Delta ID: 0.07 -> 0.55 | Improved |
| perturb only identity-related dimensions | encoder unlearning · CelebA-HQ, FastComposer, oneID | ID_target^g: 0.03 -> 0.21 | Hurt |
| perturb only identity-related dimensions | encoder unlearning · PhotoMaker transfer | Delta ID: 0.1108 -> 0.7007 | Improved |
| raise kept ratio k of identity-related dimensions (ablation) | IREU · CelebA-HQ | Delta FID | Hurt |
| raise perturbation magnitude alpha (ablation) | IREU · CelebA-HQ | ID_target similarity | Improved |
| different random swap target x_a for the mask (ablation) | IREU · CelebA-HQ | mask overlap: 70.46% +/- 5.19% | No clear effect |
| encoder-only unlearning (IREU) vs BIA on unseen images of target | IREU · FastComposer oneID to fiveID | ID_target^o,unseen: 0.32 -> 0.15 | Improved |
Closest papers by meaning
- Generative Unlearning for Any Identity (CVPR 2024 · similarity 0.87)
- IDProtector: An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation (CVPR 2025 · similarity 0.87)
- FaceA-Net: Facial Attribute-Driven ID Preserving Image Generation Network (AAAI 2025 · similarity 0.85)
- PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding (CVPR 2024 · similarity 0.85)
- Adv-CPG: A Customized Portrait Generation Framework with Facial Adversarial Attacks (CVPR 2025 · similarity 0.85)
- PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization (CVPR 2024 · similarity 0.84)
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.
