Research map / Image generation / Autoregressive and latent generation
Free-Form Image Inpainting With Gated Convolution
Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang
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 rule-based partial convolution with learnable gated convolution | inpainting network (same framework) · Places2 val (free-form mask) | mean l1 error: 10.4% -> 9.1% | Improved |
| replace rule-based partial convolution with learnable gated convolution | inpainting network (same framework) · Places2 val (free-form mask) | mean l2 error: 1.9% -> 1.6% | Improved |
| replace rule-based partial convolution with learnable gated convolution | inpainting network (same framework) · Places2 val (rectangular mask) | mean l1 error: 9.8% -> 8.6% | Improved |
| replace rule-based partial convolution with learnable gated convolution | inpainting network (same framework) · Places2 val (rectangular mask) | mean l2 error: 2.3% -> 2.0% | Improved |
| use gated-convolution network instead of reimplemented PartialConv | inpainting network · Places2 user study naturalness | naturalness rating (1-10): 7.07 -> 7.72 | Improved |
| use gated-convolution model instead of official PartialConv model | inpainting system · Places2 user study pairwise | preference (%): 79.4% vs. 20.6% | Improved |
| use gated-convolution model instead of official PartialConv model | inpainting system · Places2 user study naturalness | naturalness rating (1-10): 6.54 -> 7.72 | Improved |
| add SN-PatchGAN (spectral-normalized patch discriminator) instead of one global GAN without spectral normalization (ablation) | gated convolution inpainting network · Places2 qualitative examples | visual quality | Improved |
| add gated convolution to Global&Local-style rectangle-mask baselines on free-form masks (compare Ours vs Global&Local) | inpainting network · Places2 val (free-form mask) | mean l1 error: 21.6% -> 9.1% | Improved |
| add extra pixel-wise loss on HED output features for sketch guidance (ablation) | user-guided inpainting model · Places2 sketch-guided inpainting | inpainting quality | No clear effect |
| slim model width by 25% after adding gated convolutions (ablation) | gated convolution inpainting network · Places2 | performance: 25% | No clear effect |
| add U-Net skip connections (ablation) | encoder-decoder inpainting network · Places2 non-narrow masks | inpainting quality | No clear effect |
Closest papers by meaning
- Free-Form Video Inpainting With 3D Gated Convolution and Temporal PatchGAN (ICCV 2019 · similarity 0.90)
- Generative Image Inpainting With Contextual Attention (CVPR 2018 · similarity 0.89)
- InOut: Diverse Image Outpainting via GAN Inversion (CVPR 2022 · similarity 0.89)
- Image Inpainting via Iteratively Decoupled Probabilistic Modeling (ICLR 2024 · similarity 0.89)
- MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices (ICCV 2023 · similarity 0.88)
- Image Inpainting for Irregular Holes Using Partial Convolutions (ECCV 2018 · similarity 0.88)
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.
