Research map / Image generation / Autoregressive and latent generation
Neural FFTs for Universal Texture Image Synthesis
Morteza Mardani, Guilin Liu, Aysegul Dundar, Shiqiu Liu, Andrew Tao, Bryan Catanzaro
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 |
|---|---|---|---|
| add FFT-domain upsampling of features (vs CNN on features without FFT) (ablation) | Neural FFT · Texture test set (128->256, 200 images) | FID: 147.6 -> 71.82 | Improved |
| add FFT-domain upsampling of features (ablation) | Neural FFT · Texture test set (128->256, 200 images) | SSIM: 0.29 -> 0.48 | Improved |
| upsample in feature space instead of end-to-end FFT of the image (ablation) | Neural FFT · Texture test set (128->256, 200 images) | FID: 126.3 -> 71.82 | Improved |
| upsample in feature space instead of end-to-end FFT of the image (ablation) | Neural FFT · Texture test set (128->256, 200 images) | SSIM: 0.28 -> 0.48 | Improved |
| add deformable convolutions to the FFT upsampler (ablation) | Neural FFT · Texture test set (128->256, 200 images) | FID: 76.00 -> 71.82 | Improved |
| add deformable convolutions to the FFT upsampler (ablation) | Neural FFT · Texture test set (128->256, 200 images) | SSIM: 0.46 -> 0.48 | Improved |
| add deformable convolutions to the FFT upsampler (ablation) | Neural FFT · Texture test set (128->256, 200 images) | LPIPS: 0.27 -> 0.25 | Improved |
| Neural FFT vs Texture CNN (which uses full ground-truth texture) | Neural FFT · Texture test set (128->256, 200 images) | FID: 68.97 -> 71.82 | Hurt |
| Neural FFT vs self-tuning optimization baseline | Neural FFT · Texture test set (128->256, 200 images) | human preference / runtime: PS 58; 45 ms vs 140 s | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Texture mixer — Texture Mixer: A Network for Controllable Synthesis and Interpolation of Texture (CVPR 2019) · better on 2 of 2 shared results
- sinGAN — SinGAN: Learning a Generative Model From a Single Natural Image (ICCV 2019) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Fast Fourier Convolution in Image Inpainting (ICCV 2023 · similarity 0.87)
- Diversified Texture Synthesis With Feed-Forward Networks (CVPR 2017 · similarity 0.86)
- Deep Fourier Up-Sampling (NeurIPS 2022 · similarity 0.86)
- Learning Local Implicit Fourier Representation for Image Warping (ECCV 2022 · similarity 0.86)
- Texture Networks: Feed-forward Synthesis of Textures and Stylized Images (ICML 2016 · similarity 0.86)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.84)
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.
