Research map / Image generation / GANs
GH-Feat: Learning Versatile Generative Hierarchical Features From GANs
Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, Bolei Zhou
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 |
|---|---|---|---|
| Encode into Y space instead of W space (with SAM) (ablation) | GH-Feat · FF-HQ reconstruction | FID: 22.24 -> 18.48 | Improved |
| Add spatial alignment module (SAM) (ablation) | GH-Feat (Y space) · FF-HQ reconstruction | FID: 19.06 -> 18.48 | Improved |
| Add spatial alignment module (SAM) (ablation) | GH-Feat (Y space) · FF-HQ reconstruction | MSE: 0.0502 -> 0.0464 | Improved |
| Add regularizer (ablation) | GH-Feat · FF-HQ reconstruction | FID: 18.48 -> 16.84 | Improved |
| Add regularizer (ablation) | GH-Feat · FF-HQ reconstruction | MSE: 0.0464 -> 0.0494 | Hurt |
| Add regularizer (ablation) | GH-Feat · Cosine similarity encoder vs native latent | cosine similarity: 0.444 -> 0.879 | Improved |
| Use pre-trained StyleGAN as loss instead of training generator from scratch (ablation) | GH-Feat · FF-HQ reconstruction | FID: 46.20 -> 18.48 | Improved |
| Use spatial GH-Feat instead of plain GH-Feat | GH-Feat · FF-HQ faces reconstruction | MSE: 0.046 -> 0.029 | Improved |
| Use spatial GH-Feat instead of plain GH-Feat | GH-Feat · LSUN bedroom reconstruction | MSE: 0.068 -> 0.057 | Improved |
| Add regularizer (ablation) | GH-Feat · MNIST digit recognition | accuracy (%): 99.06 -> 98.78 | Hurt |
| Add regularizer (voting) (ablation) | GH-Feat · LFW face verification | accuracy (%): 69.7 -> 69.1 | Hurt |
| Use voting instead of grouping across levels | GH-Feat · LFW face verification | accuracy (%): 60.1 -> 69.7 | Improved |
Closest papers by meaning
- Generative Hierarchical Features From Synthesizing Images (CVPR 2021 · similarity 0.96)
- Semantic Hierarchy Emerges in Deep Generative Representations for Scene Synthesis (IJCV 2021 · similarity 0.90)
- Linear Semantics in Generative Adversarial Networks (CVPR 2021 · similarity 0.88)
- Semantic Pyramid for Image Generation (CVPR 2020 · similarity 0.88)
- Dual Contrastive Loss and Attention for GANs (ICCV 2021 · similarity 0.88)
- Local and Global GANs With Semantic-Aware Upsampling for Image Generation (TPAMI 2022 · 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.
