GNN-inpaint: Multi-Method Mural Inpainting Workspace

This repository is a single workspace that combines multiple mural inpainting pipelines:

• VAE mask generation
• PowerPaint + ControlNet
• Anyline preprocessing + MuralNet
• DeepFill
• LaMa
• Evaluation/comparison

The goal is to keep one shared data convention and evaluate all methods with the same inputs.

1. Repository Layout

GNN-inpaint/
├── Data/                               # Shared data root for all projects
│   ├── Mural_seg_processed/
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   ├── val/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   └── test/
│   │       ├── images/
│   │       └── labels/
│   ├── MuralRealistic/
│   │   ├── original_images/
│   │   ├── masks/
│   │   ├── edges/
│   │   ├── damaged_albedo_dark/
│   │   ├── white_block_damaged/
│   │   ├── anyline-damaged/
│   │   └── anyline-white/
│   └── VAE_generated_masks/
├── GNN-anyline/
├── GNN-deepfill/
├── GNN-evaluation/
├── GNN-Lama/
├── GNN-muralnet/
├── GNN-powerpaint+controlnet/
└── GNN-VAE/

2. Data Directory Explained (Whole Project)

Data/ is the shared dataset root for all subprojects. Treat it as the canonical source of truth.

2.1 Data/Mural_seg_processed/

Use for training/validation/testing where split-based data is required.

• train/images, val/images, test/images: input images.
• train/labels, val/labels, test/labels: mask/label targets used by mask-related training (for example VAE mask learning).

Used by:

• GNN-VAE/ for mask learning (VAE_learn_mask.ipynb).
• Any training scripts that need split-based supervised data.

2.2 Data/MuralRealistic/

Use for realistic restoration/inference/evaluation scenarios.

• original_images/: source mural images.
• masks/: binary masks for inpainting.
• edges/: edge/line hints.
• damaged_albedo_dark/, white_block_damaged/: damaged input variants.
• anyline-damaged/, anyline-white/: Anyline-derived guidance.

Used by:

• GNN-anyline/ preprocessing flow.
• Inference pipelines in DeepFill / MuralNet / LaMa / PowerPaint (depending on each script config).
• GNN-evaluation/ for cross-method comparison on consistent samples.

2.3 Data/VAE_generated_masks/

Use for generated masks exported from VAE and reused by other methods.

Used by:

• GNN-VAE/ as output target directory.
• Other model pipelines when synthetic mask augmentation is needed.

3. Subproject Responsibilities

• GNN-anyline/: generate/prepare line guidance and related assets.
• GNN-VAE/: train VAE mask generator and export mask sets.
• GNN-deepfill/: train/test DeepFill and produce inpainting outputs.
• GNN-muralnet/: run MuralNet training/inference and checkpoints.
• GNN-Lama/: run LaMa pipeline and metric scripts.
• GNN-powerpaint+controlnet/: run PowerPaint + ControlNet experiments.
• GNN-evaluation/: gather all methods' outputs and run comparisons.

4. End-to-End Usage (All Projects)

1. Prepare shared datasets under Data/.
2. Optionally run GNN-anyline/scripts/ to generate line guidance.
3. Optionally run GNN-VAE/ to generate extra masks and store them under Data/VAE_generated_masks/.
4. Run each inpainting method in its own folder:
   • GNN-deepfill/
   • GNN-muralnet/
   • GNN-Lama/
   • GNN-powerpaint+controlnet/
5. Copy/organize outputs into GNN-evaluation/ structure.
6. Run evaluation scripts in GNN-evaluation/scripts/.

5. Environment Management

Use separate environments per method to avoid dependency conflicts.

• GNN-anyline/.venv
• GNN-deepfill/venv (or .venv)
• GNN-muralnet/.venv
• GNN-Lama/.venv
• GNN-powerpaint+controlnet/.venv

6. GitHub Tracking and Ignore Policy

This repository should track code and configs, but not large data/results/models.

Track in Git:

• Python source code and scripts.
• Configuration files.
• Documentation and lightweight notebooks.

Ignore in Git (already covered in root .gitignore):

• Entire shared data root: Data/.
• Model/checkpoint files: *.pth, *.pt, *.ckpt, *.safetensors, *.bin, *.onnx.
• Generated outputs in GNN-evaluation/*-result/, comparison folders, and method result folders.
• Virtual environments and caches.
• Large archives (*.zip, *.rar, *.tar, *.gz, *.7z).

7. Quick Start

cd GNN-inpaint

# Example: VAE workflow
cd GNN-VAE
jupyter notebook VAE_learn_mask.ipynb

For each model, run commands from that subproject's README or entry script.