KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#developer-tools #codex #typescript #ai-agents #ai-agent #cli #claude-code #ios #open-source #deepseek-harness #dsh-plugin #python

Normal-map-generator

★ 79 Open GitHub ↗

Generate a normal map or displacement map from a photo texture with UNet

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Archive message

This repo is archived now. I don't plan on doing any major updates to it. Feel free to fork and update it!

# Normal map generator
Aside from Base Color texture images, 3D texture packs often include images such as Normal maps, Displacement maps to make the texture look more realistic. Blender Guru has an excellent video explaining what each texture map does in a 3D animation software: [Link](https://www.blenderguru.com/tutorials/basics-realistic-texturing)

![Textureing](https://images.squarespace-cdn.com/content/v1/58586fa5ebbd1a60e7d76d3e/1494407035745-K9TK8VNTL62E3J98RCU5/image-asset.jpeg?format=400w)

A texture pack from [ambientcg.com](https://ambientcg.com/) includes images below:  
| Base Color | Normal Map | Displacement Map | Roughness |
|:--:|:--:|:--:|:--:|
| ![color](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/brick_texture/color.jpg) | ![normal](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/brick_texture/normal.jpg) | ![disp](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/brick_texture/displacement.jpg) | ![rough](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/brick_texture/roughness.jpg) |

This project aims to generate normal maps and displacement maps automatically with UNet, using texture images from [ambientcg.com](https://ambientcg.com/) as our training dataset.

## Results
Training output:  
![out1](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/results/output1.png)  
![out2](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/results/output2.png)

Blender demo with custom photo texture:
| No Normal Map | With Normal Map |
|:--:|:--:|
| ![nonorm](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/results/nonormal.gif) | ![withnorm](https://github.com/yuchen071/Normal-map-generator/blob/main/.readme_docs/results/withnormal.gif) |

The `test/input/` folder should contain your own images you wish to be converted.

### Dependencies
```
torchinfo==0.1.1
matplotlib==3.3.4
numpy==1.21.0
torchvision==0.9.0
fake_useragent==0.1.11
torch==1.8.0
requests==2.26.0
tqdm==4.62.2
beautifulsoup4==4.10.0
Pillow==9.0.1
```

### Train & Validation
Run the following script to train a network that generates normal maps:
```
python train_norm.py
```
Run the following script to train a network that generates displacement maps:
```
python train_disp.py
```
A `valid/` folder should appear in root during training containing validation sample outputs.  
A `checkpoints/` folder should also appear after training script ends containing trained network `.pth` files.

#### Parameters
Global parameters can be tinkered in the script:  
```python
DIR_TRAIN         # str, training dataset folder
DIR_VALID         # str, validation folder for output
DIR_TEST          # str, testing folder with custom images
CHK_OUT           # str, model checkpoint output folder
TEST_CROP         # int [px], center crop of custom testing images

# Training parameters
PARAMS = {
    "Type":       # str, just a name

"pretrain":   # str or None, .pth filename in CHK_OUT folder to continue training, change to None to train from scratch

"train": {
        "epochs":       # int, training epochs
        "batch":        # int, batch size
        "lr":           # float, learning rate
        "split":        # float 0~1, split percentage between training and validation dataset
        "nWorkers":     # int, Dataloader worker numbers
        "log_interv":   # int, epoch interval for valid image output
        },

"image": {
        "img_resize":   # int [px], image resize size
        "img_crop":     # int [px], image center crop size
        "rand_flip":    # bool, flip training images randomly to add variation
        "rand_crop":    # int [px] or None, random crop in training image to add variation
        },

"writer": False,    # bool, Tensorboard on/off
}
```

### Evaluation
Run the following script to evaluate the trained normal map generator network:
```
python eval_norm.py
```
Run the following script to evaluate the trained displacement map generator network:
```
python eval_disp.py
```
By default it reads images from the same `test/` folder and outputs generated images into `test/output/` folder. Can be changed in global parameters.
#### Parameters
Global parameters can be tinkered in the script:  
```python
PATH_CHK    # str, path/to/network_checkpoint.pth
DIR_EVAL    # str, evaluation folder with custom images
CROP        # int [px], image resize and center crop size
```

### How to import and use in Blender
Refer to the tutorial from the ambientcg help page: [Link](https://docs.ambientcg.com/books/using-the-assets/page/pbr-materials-in-blender)

Normal-map-generator

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README