---
license: mit
library_name: diffusers
base_model: timbrooks/instruct-pix2pix
datasets:
- sttkw/TAID-Dataset
- sttkw/TAID-AtmosEdit
tags:
- intrinsic-decomposition
- terrain
- atmosphere
---

# TAID Models

Pretrained weights for *Atmosphere-Aware Intrinsic Decomposition from a Single
Terrain Image with Latent Diffusion Models*.

- Code: <https://github.com/sttkw/terrain-atmospheric-intrinsic-decomposition>
- Project page: <https://sttkw.github.io/terrain-atmospheric-intrinsic-decomposition/>

| Path | Description |
| --- | --- |
| `decomposition/unet/` | Intrinsic decomposition U-Net (InstructPix2Pix fine-tune, step 18,000). Predicts Albedo, Diffuse Shading, Specular Shading and Volume. |
| `decomposition/terrain_decomposition_config.json` | Inference metadata (resolution, target encodings). |
| `atmosphere/terrain_difference.pt` | Atmospheric editor (12ch → 9ch U-Net) that changes D, S and V for new air / aerosol / ozone densities. |

The decomposition U-Net is loaded on top of the other components of
`timbrooks/instruct-pix2pix` (VAE, text encoder, tokenizer, scheduler).

## Usage

```bash
git clone https://github.com/sttkw/terrain-atmospheric-intrinsic-decomposition
cd terrain-atmospheric-intrinsic-decomposition
pip install -r requirements.txt
python demo.py --input_image demo_image/test1.jpg --water_mask demo_image/test1.png \
  --p_control 0 -3 0 --output_dir outputs/demo/test1
```

`demo.py` downloads these weights automatically.

## License

MIT, same as the code.