Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Prior Feature and Attention Enhanced Image Inpainting
Chenjie Cao, Qiaole Dong, Yanwei Fu
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 MAE prior features with naive element-wise addition (ablation) | LaMa-style ACR · Places2 5-scene subset 256x256 | FID: 28.01 -> 28.01 | No clear effect |
| multiply MAE prior features by a zero-initialized learnable alpha before addition (ablation) | ACR + MAE · Places2 5-scene subset 256x256 | FID: 28.01 -> 26.83 | Improved |
| initialize alpha with 1 instead of 0 (ablation) | ACR + MAE · Places2 5-scene subset 256x256 | FID: 26.83 -> 26.90 | No clear effect |
| add trainable contextual attention (ablation) | ACR + MAE · Places2 5-scene subset 256x256 | FID: 26.84 -> 26.99 | Hurt |
| add prior attention aggregation from MAE decoder (ablation) | ACR + MAE (mixed masks) · Places2 5-scene subset 256x256 | FID: 26.84 -> 25.49 | Improved |
| pre-train MAE with mixed continuous masks instead of random masks (ablation) | FAR with prior attention · Places2 5-scene subset 256x256 | FID: 26.25 -> 25.49 | Improved |
| finetune MAE for partially masked patches (ablation) | FAR full · Places2 5-scene subset 256x256 | FID: 25.49 -> 26.00 | Hurt |
| add 2D coordinate grid to resized prior features (ablation) | FAR · Places2 512x512 finetuned | FID: 25.47 -> 25.33 | Improved |
| remove patch-wise pixel normalization from MAE target (ablation) | MAE pre-training · Places2 512x512 | quality: 25.33 -> 25.30 | Improved |
| replace LaMa with FAR on Places2 256 | LaMa · Places2 256x256 | FID: 1.63 -> 1.31 | Improved |
| replace LaMa with FAR on Places2 512 | LaMa · Places2 512x512 | FID: 27.8 -> 25.3 | Improved |
| replace LaMa with FAR on FFHQ 256 | LaMa · FFHQ 256x256 | FID: 6.38 -> 6.15 | Improved |
Closest papers by meaning
- Incremental Transformer Structure Enhanced Image Inpainting With Masking Positional Encoding (CVPR 2022 · similarity 0.89)
- Image Inpainting With Learnable Bidirectional Attention Maps (ICCV 2019 · similarity 0.88)
- Localization of Deep Inpainting Using High-Pass Fully Convolutional Network (ICCV 2019 · similarity 0.88)
- Deep Learning-Based Image and Video Inpainting: A Survey (IJCV 2024 · similarity 0.88)
- Prior Guided GAN Based Semantic Inpainting (CVPR 2020 · similarity 0.88)
- Distillation-Guided Image Inpainting (ICCV 2021 · 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.
