Research map / Image generation / Autoregressive and latent generation
Image Generators With Conditionally-Independent Pixel Synthesis
Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, Denis Korzhenkov
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 |
|---|---|---|---|
| replace spatial-convolution synthesis with conditionally independent coordinate-based pixel synthesis | StyleGANv2-style image generation · Multiple datasets at 256×256 | FID: FFHQ: 3.83 -> 4.38; LSUN Churches: 3.86 -> 2.92; Landscapes: 2.36 -> 3.61; Satellite-Buildings: 73.67 -> 69.67; Satellit | Mixed |
| replace spatial-convolution synthesis with conditionally independent coordinate-based pixel synthesis | StyleGANv2 · FFHQ 256×256 | Precision: 0.609 -> 0.613 | Improved |
| replace spatial-convolution synthesis with conditionally independent coordinate-based pixel synthesis | StyleGANv2 · FFHQ 256×256 | Recall: 0.513 -> 0.493 | Hurt |
| add coordinate embeddings (ablation) | CIPS-base · FFHQ | FID: 12.71 -> 6.71 | Improved |
| add Fourier features (ablation) | CIPS-No Fourier · FFHQ | FID: 10.18 -> 6.31 | Improved |
| add cumulative skip connections to RGB outputs (ablation) | CIPS-base · FFHQ | FID: 6.71 -> 6.31 | Improved |
| use residual blocks instead of cumulative skip projections to RGB (ablation) | CIPS · FFHQ | FID: 6.31 -> 6.52 | Hurt |
| use sine activation in all layers (ablation) | CIPS · FFHQ | FID: 6.31 -> 10.0 | Hurt |
| reduce training patch size from 256 to 128 (ablation) | CIPS · FFHQ | FID: 4.38 -> 9.08 | Hurt |
| reduce training patch size from 256 to 64 (ablation) | CIPS · FFHQ | FID: 4.38 -> 11.79 | Hurt |
| reduce training patch size from 256 to 128 (ablation) | CIPS · LSUN Churches | FID: 2.92 -> 7.08 | Hurt |
| reduce training patch size from 256 to 64 (ablation) | CIPS · LSUN Churches | FID: 2.92 -> 11.53 | Hurt |
Closest papers by meaning
- Spatially-Adaptive Pixelwise Networks for Fast Image Translation (CVPR 2021 · similarity 0.88)
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (ICLR 2019 · similarity 0.88)
- PixelNN: Example-based Image Synthesis (ICLR 2018 · similarity 0.87)
- Cluster-Guided Image Synthesis With Unconditional Models (CVPR 2022 · similarity 0.87)
- Semi-supervised FusedGAN for Conditional Image Generation (ECCV 2018 · similarity 0.87)
- Unlocking Pre-trained Image Backbones for Semantic Image Synthesis (CVPR 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.
