This project explores how modern generative AI transforms image inpainting compared to classic techniques. By benchmarking two state‐of‐the‐art generative models—LaMa and Stable Diffusion—against the traditional Telea method, the study reveals significant improvements in both performance and visual quality.
- Python
- PyTorch
- Stable Diffusion
- LaMa
- OpenCV
- Pandas
- Image preparation
- Mask improvement
- Model evaluation
- Bootstrapping
- Significance testing
- LaMa
- Stable Diffusion (Hugging Face)
- Telea (OpenCV)
- Superior Performance: AI-based models (LaMa & Stable Diffusion) outperform Telea by approximately 14%.
- Mask Size is Critical: The missing area's size is the most influential factor affecting inpainting quality.
- Model Suitability:
- Stable Diffusion excels with large masks (though it may sometimes generate "hallucinations").
- LaMa is optimal for small-to-medium masks but struggles with larger missing areas.
- Future Directions: Enhancing model efficiency and reducing unwanted object generation will further improve inpainting quality.
Image inpainting is the process of restoring missing, damaged, or removed parts of an image. It is widely used for image restoration, enhancement, and object removal.
| Model | Improvement to Baseline (%) | Ranking |
|---|---|---|
| LaMa | +0.17% | 1 |
| Telea (Baseline) | 0% | 2 |
| Stable Diffusion | -21.48% | 3* |
| Model | Improvement to Baseline (%) | Ranking |
|---|---|---|
| LaMa | +11.67% | 1* |
| Telea (Baseline) | 0% | 3* |
| Stable Diffusion | +3.83% | 2* |
| Model | Improvement to Baseline (%) | Ranking |
|---|---|---|
| LaMa | +14.27% | 2* |
| Telea (Baseline) | 0% | 3* |
| Stable Diffusion | +18.78% | 1* |
| Model | Improvement to Baseline (%) | Ranking |
|---|---|---|
| LaMa | +14.64% | 1* |
| Telea (Baseline) | 0% | 3* |
| Stable Diffusion | +13.40% | 2* |