### Features
- **Cross-ModalSemantic Inpainting with Robust Occlusion-Free Support (robustness)**  
  Restore masked or occluded regions using self-supervised cross-modal context.

- **Flexible Design with Optional Pretrained Architectures for Multiple Downstream Tasks (generalibility)**  
  Semantic Encoder(s) → Semantic Inpaintor (Projection) → Downstream Decoder(s)
  <br>Reuse components across **versatile tasks**, here we give three examples:
  - Image Inpainting  
  - Future CSI Prediction  
  - Human Count Estimation
  
- Compress

---

## Project Structure (in progress...)
```text
semantic-inpainting-framework/
├── code/                     #  Encoders, inpaintors, and decoders
|     ├── Step 3: Pretraining for Downstream Decoders
|     ├── Step 2: Pretraining for Semantic Inpaintors (projection heads)
|     └── Step 1: Pretraining for Semantic Encoders
├── data/                     # Pre-/post-processing scripts
├── experiments/              # Training & evaluation pipelines
├── pretrained models/        # Pretrained components
├── Images/            　　　　# Figures for docs
└── README.md