Research map / Image generation / GANs
Positional Encoding As Spatial Inductive Bias in GANs
Rui Xu, Xintao Wang, Kai Chen, Bolei Zhou, Chen Change Loy
Check 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 |
|---|---|---|---|
| Remove zero padding (ablation) | StyleGAN2 · FFHQ | FID@256, 20M: 5.79 -> 6.45 | Hurt |
| Remove zero padding (ablation) | PGGAN · CelebA cropped 128x128 | SWD Avg. (× 10^3): 4.149 -> 8.534 | Hurt |
| Add SPE to the padding-free PGGAN (ablation) | PGGAN · CelebA cropped 128x128 | SWD Avg. (× 10^3): 8.534 -> 5.694 | Improved |
| Remove zero padding (ablation) | DCGAN · LSUN Bedroom 128x128 | SWD Avg. (×10^3): 19.07 -> 31.24 | Hurt |
| Add SPE to the padding-free DCGAN (ablation) | DCGAN · LSUN Bedroom 128x128 | SWD Avg. (×10^3): 31.24 -> 17.93 | Improved |
| Add SPE to the padding-free PGGAN (ablation) | PGGAN · LSUN Bedroom 128x128 | SWD Avg. (× 10^3): 10.081 -> 5.371 | Improved |
| Use expanded SPE in multi-scale training | StyleGAN2-C2-256 · FFHQ | FID@512, 25M: 3.41 -> 2.93 | Improved |
| Use expanded rather than interpolated SPE (ablation) | MS-PIE StyleGAN2-C2-256 · FFHQ | FID@512, 25M: 3.15 -> 2.93 | Improved |
| Use Cartesian grid rather than expanded SPE (ablation) | MS-PIE StyleGAN2-C2-256 · FFHQ | FID@512, 25M: 2.93 -> 3.50 | Hurt |
| Reduce channel multiplier from two to one (ablation) | MS-PIE SPE-expand · FFHQ | FID@512, 25M: 2.93 -> 3.40 | Hurt |
| Use a learnable constant input instead of expanded SPE (ablation) | MS-PIE StyleGAN2-C2-256 · FFHQ | FID@512, 25M: 2.93 -> 3.35 | Hurt |
| Use padding-free training rather than retaining zero padding (ablation) | MS-PIE StyleGAN2-C2-256 · FFHQ | FID@512, 25M: 2.93 -> 4.17 | Hurt |
Closest papers by meaning
- Toward Spatially Unbiased Generative Models (ICCV 2021 · similarity 0.88)
- StyleSwin: Transformer-Based GAN for High-Resolution Image Generation (CVPR 2022 · similarity 0.87)
- SinGAN: Learning a Generative Model From a Single Natural Image (ICCV 2019 · similarity 0.87)
- Spatial Frequency Bias in Convolutional Generative Adversarial Networks (AAAI 2022 · similarity 0.86)
- Spatial Steerability of GANs via Self-Supervision from Discriminator (TPAMI 2024 · similarity 0.86)
- WeditGAN: Few-Shot Image Generation via Latent Space Relocation (AAAI 2024 · similarity 0.85)
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.
