Research map / Image generation / Editing and style transfer
Gradual Residuals Alignment: A Dual-Stream Framework for GAN Inversion and Image 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 GradStyle, a dual-stream StyleGAN2 inversion and attribute-editing framework that gradually injects coarse-to-fine residual details at multiple generator stages, with selective attention to align them to edited features.
Problem
GAN inversion for attribute editing trades off detail preservation against editability; injecting all high-bit lost details at once causes misaligned details, inconsistent content, and artifacts in edited images.
Approach
Inject lost details gradually in a multi-stage coarse-to-fine manner in both reconstruction and editing: a Gradual Residual Module computes gated residual features per stage, and a Global Alignment Module uses Selective Attention (top-μ% masked attention) to align these residuals to the edited feature maps, trained self-supervised via random perspective transforms.
Key results
- On CelebA-HQ reconstruction, GradStyle reaches ID 0.813, LPIPS 0.269 and PSNR 24.583, beating CLCAE (ID 0.719, PSNR 24.375) and StyleRes (ID 0.758).
- Pose(+) editing ID is 0.677 for GradStyle versus 0.675 for CLCAE; smile(+) ID is 0.690 versus 0.653 for CLCAE.
- Ablation: removing GAM drops editing ID from 0.677 to 0.496 and removing GRM drops it to 0.543.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GradStyle instead of CLCAE | StyleGAN2 inversion · Celeba-HQ reconstruction | ID: 0.719 -> 0.813 | Helped |
| use GradStyle instead of CLCAE | StyleGAN2 inversion · Celeba-HQ reconstruction | LPIPS: 0.289 -> 0.269 | Helped |
| use GradStyle instead of CLCAE | StyleGAN2 attribute editing · Pose(+) editing | ID: 0.675 -> 0.677 | Helped |
| remove GAM (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.708 | Hurt |
| remove GRM (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.583 | Hurt |
| remove both GAM and GRM (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.509 | Hurt |
| add residuals at one low-level stage only (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.504 | Hurt |
| add residuals at one high-level stage only (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.763 | Hurt |
| use hierarchical encoder feature I only (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.702 | Hurt |
| use hierarchical encoder feature II only (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.742 | Hurt |
| use hierarchical encoder feature III only (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.785 | Hurt |
| add refinement at the 10th generator block (ablation) | GradStyle · Celeba-HQ reconstruction | ID: 0.813 -> 0.873 | Helped |
Surprising findings
- Adding 64x64 refinement (10th block) raises reconstruction ID to 0.873 but lowers pose-editing ID to 0.636, so more refinement stages do not always help editing.
- Dense refinement of blocks 3-8 gives reconstruction ID 0.816 and pose ID 0.665, no better than the chosen 3-stage setup.
Limitations
- Struggles with wrinkle addition and large color changes such as pale hair editing, since random perspective transforms do not simulate these.
- Editing evaluated mainly with ID score; other editing metrics deemed unsuitable by the authors.
- Main experiments limited to face domain with two editors (InterfaceGAN, GANspace).
Experimental setup
Datasets: FFHQ CelebA-HQ Stanford Cars MetFaces
Benchmarks: CelebA-HQ
Metrics: ID SSIM L2 LPIPS PSNR
Baselines: e4e ReStyle HyperStyle HFGI StyleRes CLCAE
Builds on: StyleGAN2 e4e HFGI StyleRes InterfaceGAN
Closest papers by meaning
- High-Fidelity GAN Inversion for Image Attribute Editing (CVPR 2022 · similarity 0.93)
- The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing (CVPR 2024 · similarity 0.88)
- Spatial-Contextual Discrepancy Information Compensation for GAN Inversion (AAAI 2024 · similarity 0.87)
- In-N-Out: Faithful 3D GAN Inversion with Volumetric Decomposition for Face Editing (CVPR 2024 · similarity 0.87)
- Tuning-Free Inversion-Enhanced Control for Consistent Image Editing (AAAI 2024 · similarity 0.86)
- WarpGAN: Warping-Guided 3D GAN Inversion with Style-Based Novel View Inpainting (NeurIPS 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.