Research map / Label-efficient and robust learning / Knowledge distillation
Mind the Gap in Distilling StyleGANs
Guodong Xu, Yuenan Hou, Ziwei Liu, Chen Change Loy
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 |
|---|---|---|---|
| inherit teacher mapping network weights (ablation) | StyleGAN2 student distillation · FFHQ 256x256 StyleGAN2 ablation | FID: 11.27 -> 8.23 | Improved |
| inherit teacher convolution weights in addition to mapping network (ablation) | StyleGAN2 student distillation · FFHQ 256x256 StyleGAN2 ablation | FID: 7.94 -> 8.45 | Hurt |
| add RGB+LPIPS mimicking when mapping network is randomly initialized (ablation) | StyleGAN2 student distillation · FFHQ 256x256 StyleGAN2 ablation | FID: 10.92 -> 11.27 | Hurt |
| add RGB+LPIPS mimicking when mapping network is inherited (ablation) | StyleGAN2 student distillation · FFHQ 256x256 StyleGAN2 ablation | FID: 10.54 -> 8.61 | Improved |
| add latent-direction (LD) relation loss (ablation) | StyleGAN2 student (inherit mapping) · FFHQ 256x256 StyleGAN2 ablation | FID: 8.23 -> 7.94 | Improved |
| inherit mapping network instead of random init (ablation) | StyleGAN2 student (random conv) · FFHQ 256x256 mapping network study | FID: 11.78 -> 8.30 | Improved |
| two-stage mapping mimicking with L1 instead of random init (ablation) | StyleGAN2 student · FFHQ 256x256 mapping network study | FID: 11.78 -> 9.69 | Improved |
| use L1 instead of L2 for mapping mimicking (ablation) | two-stage mapping training · FFHQ 256x256 mapping network study | FID: 10.80 -> 9.69 | Improved |
| use latent-direction augmentation instead of random offset (ablation) | relation distillation · FFHQ 256x256 relation ablation | FID: 9.80 -> 9.05 | Improved |
| use relation between normal and LD-augmented samples instead of single view (ablation) | relation distillation · FFHQ 256x256 relation ablation | FID: 9.47 -> 9.05 | Improved |
| use KL instead of L2 for LD loss with RGB+LPIPS (ablation) | LD loss · FFHQ 256x256 relation ablation | FID: 8.64 -> 8.26 | Improved |
| use proposed method instead of CAGAN | StyleGAN2 distillation · FFHQ 256x256 | FID: 7.9 -> 7.25 | Improved |
Closest papers by meaning
- Analyzing and Improving the Image Quality of StyleGAN (CVPR 2020 · similarity 0.91)
- StyleGAN2 Distillation for Feed-forward Image Manipulation (ECCV 2020 · similarity 0.90)
- Content-Aware GAN Compression (CVPR 2021 · similarity 0.90)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.87)
- Adversarial Score Distillation: When score distillation meets GAN (CVPR 2024 · similarity 0.87)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.87)
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.
