Research map / Image generation / GANs
Toward Spatially Unbiased Generative Models
Jooyoung Choi, Jungbeom Lee, Yonghyun Jeong, Sungroh Yoon
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 |
|---|---|---|---|
| use multi-scale explicit positional encoding instead of constant tensor | StyleGAN2 · FFHQ 512^2 from 256^2 model | FID: 154.40 -> 30.41 | Improved |
| use multi-scale explicit positional encoding instead of constant tensor | StyleGAN2 · FFHQ 384^2 from 256^2 model | FID: 43.11 -> 16.03 | Improved |
| use multi-scale instead of single-scale positional encoding (SS-PE) (ablation) | StyleGAN2 with explicit PE · FFHQ 384^2 from 256^2 model | FID: 94.44 -> 16.03 | Improved |
| use multi-scale instead of single-scale positional encoding (SS-PE) (ablation) | StyleGAN2 with explicit PE · FFHQ 256^2 | FID: 7.37 -> 6.75 | Improved |
| add MS-PE to StyleGAN2 | StyleGAN2 · FFHQ 256^2 | FID: 7.21 -> 6.75 | Improved |
| add MS-PE to StyleGAN2 | StyleGAN2 · FFHQ 256^2 | Recall: 35.20 -> 42.27 | Improved |
| add MS-PE to StyleGAN2 | StyleGAN2 · FFHQ 256^2 | Precision: 64.07 -> 62.97 | Hurt |
| add MS-PE to StyleGAN2 | StyleGAN2 · LSUN-Church 256^2 | FID: 4.70 -> 6.42 | Hurt |
| add MS-PE to StyleGAN2 | StyleGAN2 · LSUN-Church 256^2 | Recall: 39.74 -> 34.20 | Hurt |
| add random resizing of positional encoding with resized generator outputs (ablation) | MS-PE · FFHQ multi-scale generation | sample quality | Improved |
| add MS-PE to DDPM trained on color-MNIST biased to upper-left | DDPM · color-MNIST reconstruction from bottom-left input | qualitative position fidelity | Improved |
Closest papers by meaning
- Positional Encoding As Spatial Inductive Bias in GANs (CVPR 2021 · similarity 0.88)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.87)
- Understanding the Generalization of Pretrained Diffusion Models on Out-of-Distribution Data (AAAI 2024 · similarity 0.87)
- RA-ClipScore: Making Generative Model Evaluation More Interpretable (NeurIPS 2026 · similarity 0.86)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.85)
- Distance Marching for Generative Modeling (NeurIPS 2026 · 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.
