Research map / Image generation / Editing and style transfer
SemanticStyleGAN: Learning Compositional Generative Priors for Controllable Image Synthesis and Editing
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace SemanticGAN's patch discriminator with the proposed dual-branch discriminator (ablation) | SemanticGAN training on CelebAMask-HQ · CelebAMask-HQ 256×256 | FID: 18.54 -> 7.50 | Helped |
| replace SemanticGAN's patch discriminator with the proposed dual-branch discriminator (ablation) | SemanticGAN training on CelebAMask-HQ · CelebAMask-HQ 256×256 | IS: 2.77 -> 3.51 | Helped |
| replace the SemanticGAN generator with compositional semantic-area local generators (ablation) | SemanticGAN + proposed training · CelebAMask-HQ 256×256 | FID: 7.50 -> 6.42 | Hurt |
| replace the SemanticGAN generator with compositional semantic-area local generators (ablation) | SemanticGAN + proposed training · CelebAMask-HQ 256×256 | IS: 3.51 -> 3.21 | Hurt |
| use joint image-and-segmentation training instead of image-only training | SemanticGAN · CelebAMask-HQ 256×256 | FID: 4.45 -> 18.54 | Hurt |
| use joint image-and-segmentation training instead of image-only training | SemanticGAN · CelebAMask-HQ 256×256 | IS: 3.40 -> 2.77 | Hurt |
| use SemanticStyleGAN rather than StyleGAN2 trained on the same CelebAMask-HQ split | Restyle-psp image reconstruction · CelebAMask-HQ test images | MSE: 0.029± 0.016 -> 0.031± 0.017 | Mixed |
| use SemanticStyleGAN rather than StyleGAN2 trained on the same CelebAMask-HQ split | Restyle-psp image reconstruction · CelebAMask-HQ test images | ID: 0.575± 0.119 -> 0.602± 0.122 | Helped |
| use SemanticStyleGAN rather than StyleGAN2 trained on the same CelebAMask-HQ split | Restyle-psp image reconstruction · CelebAMask-HQ test images | LPIPS: 0.330±0.052 -> 0.335±0.051 | Hurt |
| fine-tune SemanticStyleGAN on FFHQ for 1,000 steps (ablation) | SemanticStyleGAN trained on CelebAMask-HQ · FFHQ face synthesis and editing | diversity | Helped |
| fine-tune SemanticStyleGAN on image-only target-domain data (ablation) | SemanticStyleGAN · Toonify, MetFaces, and BitMoji | local controllability | No clear effect |
| use the corresponding semantic-area latent codes to edit a local part | SemanticStyleGAN · CelebAMask-HQ face and hair interpolation | local disentanglement | Helped |
Closest papers by meaning
- Editing in Style: Uncovering the Local Semantics of GANs (CVPR 2020 · similarity 0.93)
- Navigating the GAN Parameter Space for Semantic Image Editing (CVPR 2021 · similarity 0.90)
- Linear Semantics in Generative Adversarial Networks (CVPR 2021 · similarity 0.90)
- EditGAN: High-Precision Semantic Image Editing (NeurIPS 2021 · similarity 0.89)
- StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation (CVPR 2021 · similarity 0.89)
- The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing (CVPR 2024 · similarity 0.88)
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.