Developers

Details for developers.

Texture file structure

ML model training data

data
|__ AssetName001
    |__ color.jpg
    |__ displacement.jpg
    |__ roughness.jpg
...

Global catalog

e.g. /path/to/Catalog/

data
|__ AssetName001
    |__ 1         # 1K resolution
        |__ AssetName001_1K-JPG_Color.jpg
        ...
    ...

Project catalog

e.g. /project/Textures/

data
|__ AssetName001_1K
    |__ AssetName001_1K-JPG_Color.jpg
    ...

Diffusion model

Config file

The config file is a json file:

{
    "data": "/path/to/data",
    "results": "/path/to/results",

    # Empty string means no resume
    "resume": "/path/to/resume.pt",

    # Image resolution
    "resolution": 1024,

    "unet_dim": 16,
    "unet_dim_mults": [1, 2, 4, 8],
    "timesteps": 100,

    "epochs": 100,
    "test_interval": 10,
    "batch_size": 4,
    "grad_accum": 1,
    # Number of times train set is repeated per epoch
    "train_duplicity": 1,
    "lr_start": 1e-3,
    "lr_end": 1e-6,
}