Research map / Image generation / Editing and style transfer
SDGAN: Disentangling Semantic Manipulation for Facial Attribute Editing
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 SDGAN, a GAN for facial attribute editing that uses attribute-specific editing modules with softmax-normalized semantic masks and a mask alignment loss, supporting latent- and reference-guided style manipulation on CelebA-HQ.
Problem
Facial attribute editing struggles to precisely manipulate a target attribute while preserving unrelated attributes and identity, and to control attribute styles, due to correlations between attributes and between attributes and identity.
Approach
A semantic disentanglement generator assigns features to per-attribute editing modules (AdaIN-modulated) whose outputs are gated by softmax-normalized semantic masks, plus a semantic mask alignment strategy that uses an attribute classifier loss so masks focus on regions that support correct attribute prediction, requiring only attribute labels.
Key results
- Latent-guided: average FID 45.69 vs 53.76 for HiSD (decrease of 8.07) with average Acc 95.54%.
- Reference-guided: average FID 45.37 vs 50.71 for HiSD (decrease of 5.34) with average Acc 89.70%.
- Compared with HFGI and StyleRes on eyeglasses, SDGAN achieves lower FID and higher accuracy in the latent-guided setting (FID 49.39 vs 80.64 for StyleRes on +Eyeglasses).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SDGAN instead of HiSD | latent-guided facial attribute editing · CelebA-HQ test, average across reported attribute edits | FID: 53.76 -> 45.69 | Helped |
| Use SDGAN instead of HiSD | reference-guided facial attribute editing · CelebA-HQ test, average across reported attribute edits | FID: 50.71 -> 45.37 | Helped |
| Use SDGAN instead of StyleRes | latent-guided eyeglasses editing · CelebA-HQ test, + Eyeglasses | FID: 80.64 -> 49.39 | Helped |
| Use SDGAN instead of StyleRes | latent-guided eyeglasses editing · CelebA-HQ test, + Eyeglasses | Acc: 79.17 -> 98.52 | Helped |
| Add attribute-specific editing modules instead of one unified editing module (ablation) | SDGAN · CelebA-HQ test, + Bangs, latent-guided | FID: 18.41 -> 15.69 | Helped |
| Add semantic masks (ablation) | SDGAN · CelebA-HQ test, + Bangs, latent-guided | FID: 18.43 -> 15.69 | Helped |
| Add semantic mask alignment strategy (ablation) | SDGAN · CelebA-HQ test, + Bangs, latent-guided | FID: 17.26 -> 15.69 | Helped |
Limitations
- Evaluated only on three attributes (bangs, eyeglasses, hair color) on CelebA-HQ
- VecGAN results reported only quantitatively because its code is unavailable
- Attribute accuracy depends on a classifier with 95.0% test accuracy
Experimental setup
Datasets: CelebA-HQ
Benchmarks: CelebA-HQ
Metrics: FID attribute editing accuracy
Baselines: SDIT ELEGANT HiSD VecGAN HFGI StyleRes
Builds on: HiSD StarGANv2 SDIT VecGAN
Closest papers by meaning
- L2M-GAN: Learning To Manipulate Latent Space Semantics for Facial Attribute Editing (CVPR 2021 · similarity 0.92)
- Semantic Component Decomposition for Face Attribute Manipulation (CVPR 2019 · similarity 0.90)
- MaskGAN: Towards Diverse and Interactive Facial Image Manipulation (CVPR 2020 · similarity 0.89)
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.89)
- CA-Edit: Causality-Aware Condition Adapter for High-Fidelity Local Facial Attribute Editing (AAAI 2025 · similarity 0.86)
- Enhancing Identity-Deformation Disentanglement in StyleGAN for One-Shot Face Video Re-Enactment (AAAI 2025 · similarity 0.86)
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.