Research map / Image generation / Editing and style transfer
Black Hole-Driven Identity Absorbing in Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BIA, a black-hole-inspired method that unlearns a specific face identity in the h-space of a pre-trained diffusion model by wrapping its latent region with nearby identity latents plus a semantic loss.
Problem
Identity unlearning in diffusion models for privacy: suppressing generation of a specified face identity while preserving other attributes and image quality, where random or orthogonal latent traversals produce unrealistic outputs.
Approach
Treats the target identity's h-space region as a "black hole" bounded by an SVM hyperplane over identity-labeled samples, then pulls the k nearest non-identity latent points into it using an identity loss, perceptual loss, L2 loss and a semantic loss that keeps non-identity attributes.
Key results
- BIA achieves the lowest ID similarity (0.01 on ArcFace-based ID metric) with LPIPS 0.03 and highest SSIM 0.83 in Table 1 comparisons.
- In a user study with 100 participants over 10 identities, BIA reached 86% ID dissimilarity and 98% attribute quality satisfaction.
- GUIDE attains an attribute score of 50.0 due to near-identical generations, whereas BIA attains gender 87.8, age 82.4, expression 90.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic loss to the wrapper-loss baseline (ablation) | BIA on CelebA-HQ · CelebA-HQ | ID similarity (CircularFace): 0.08 -> 0.01 | Helped |
| add semantic loss to the wrapper-loss baseline (ablation) | BIA on CelebA-HQ · CelebA-HQ | SSIM: 0.62 -> 0.73 | Helped |
| use k=1 nearest neighbor instead of k=5 for wrapping (ablation) | BIA on CelebA-HQ · CelebA-HQ | ID similarity (CircularFace): 0.01 -> 0.57 | Hurt |
| remove identity loss from the wrapper loss (ablation) | BIA on CelebA-HQ · CelebA-HQ | identity removal | Hurt |
| remove perceptual loss from the wrapper loss (ablation) | BIA on CelebA-HQ · CelebA-HQ | realism | Hurt |
| remove L2 loss from the wrapper loss (ablation) | BIA on CelebA-HQ · CelebA-HQ | structural quality | Hurt |
| increase k for black-hole wrapping (ablation) | BIA on CelebA-HQ · CelebA-HQ | identity removal | Helped |
Surprising findings
- BIA closest (k=1) yields much higher ID similarity values (0.57-0.68), indicating a single nearest neighbor is insufficient for unlearning.
Limitations
- Continuity of latent space may cause unintended effects on nearby identities.
- Evaluation limited to CelebA-HQ faces at 256x256 resolution.
- Some metrics are mixed; the Baseline and BIA differ notably in FID and SSIM trade-offs.
Experimental setup
Datasets: CelebA-HQ
Benchmarks: CelebA-HQ
Metrics: ID similarity (CircularFace, CosFace, FaceNet) FID LPIPS SSIM attribute preservation user study ID dissimilarity attribute quality
Baselines: GUIDE Random Baseline (wrapper loss only) BIA closest (k=1)
Builds on: GUIDE h-space semantic editing (Kwon et al.) iDDPM ArcFace
Closest papers by meaning
- When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · similarity 0.87)
- Interpreting the Weight Space of Customized Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Variance as a Catalyst: Efficient and Transferable Semantic Erasure Adversarial Attack for Customized Diffusion Models (ICML 2025 · similarity 0.85)
- Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.84)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.84)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 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.