Research map / Image generation / Image restoration
Rethinking Fast Fourier Convolution in Image Inpainting
Tianyi Chu, Jiafu Chen, Jiakai Sun, Shuobin Lian, Zhizhong Wang, Zhiwen Zuo, Lei Zhao, Wei Xing, Dongming Lu
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 |
|---|---|---|---|
| replace vanilla FFC (LaMa) with UFFC module | LaMa inpainting · DTD (mask <50%) | FID: 87.79 -> 52.78 | Improved |
| add range transform and inverse transform around frequency activation (ablation) | UFFC · module ablation | PSNR: 15.32 -> 18.27 | Improved |
| add AdaClip (ablation) | UFFC · module ablation | training stability: collapsed -> 18.27 PSNR | Improved |
| add dynamic skip connection (ablation) | UFFC · module ablation | training stability: collapsed -> 18.27 PSNR | Improved |
| add absolute position embedding (ablation) | UFFC · module ablation | PSNR: 17.89 -> 18.27 | Improved |
| add statistical equalization post-processing (ablation) | UFFC · module ablation | PSNR: 18.21 -> 18.27 | Improved |
| add MAE perceptual loss to ResNet and self-perceptual losses (ablation) | UFFC · perceptual loss ablation | FID: 55.65 -> 55.16 | Improved |
| use self-perceptual loss alone (ablation) | UFFC · perceptual loss ablation | FID: 70.04 vs 55.16 with all three | Hurt |
| compare UFFC with 9x-inflated FFC and 50x longer training (ablation) | FFC · DTD | visual quality | Improved |
| UFFC vs Co-Mod and MAT on large masks | UFFC · Places2/CelebA mask >50% | FID/U-IDS: CelebA U-IDS 20.32 vs MAT 21.75 | Mixed |
Closest papers by meaning
- Fast Fourier Convolution (NeurIPS 2020 · similarity 0.88)
- Neural FFTs for Universal Texture Image Synthesis (NeurIPS 2020 · similarity 0.87)
- Beyond Spatial Domain: Cross-domain Promoted Fourier Convolution Helps Single Image Dehazing (AAAI 2025 · similarity 0.84)
- FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining (ICML 2025 · similarity 0.84)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.84)
- Video Diffusion Models Are Strong Video Inpainter (AAAI 2025 · 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.
