KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #developer-tools #ai-agent #claude-code #codex #typescript #ai-agents #mcp #cli #rust #deepseek-harness

Chehara-GAN

★ 147 Open GitHub ↗

Simple चेहरा(face) restoration

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# चेहरा-GAN

**Beta model, might have issues.**

**Demo v0.2:** [<img src="https://colab.research.google.com/assets/colab-badge.svg" align="center">](https://colab.research.google.com/github/vijishmadhavan/Chehara-GAN/blob/master/%E0%A4%9A%E0%A5%87%E0%A4%B9%E0%A4%B0%E0%A4%BE_GAN.ipynb)

**Demo v0.1:** [<img src="https://colab.research.google.com/assets/colab-badge.svg" align="center">](https://colab.research.google.com/github/vijishmadhavan/Chehara-GAN/blob/master/%E0%A4%9A%E0%A5%87%E0%A4%B9%E0%A4%B0%E0%A4%BE_GAN_v1.ipynb)

If you like what I'm doing you can:

- Follow me on [twitter](https://twitter.com/Vijish68859437)
- Check my other projects [GitHub](https://github.com/vijishmadhavan)
- You can sponsor me to support my open source work 💖 [sponsor](https://github.com/sponsors/vijishmadhavan?o=sd&sc=t)

**Changelog**
* 2022-09-010 Tested it on 64×64 -> 512×512 on FFHQ-CelebaHQ - [Example Images](#Example-Images)

* 2022-08-026 Use this [weights](https://www.dropbox.com/s/1818xtxblsyrv45/c2.pth?dl=1) to train it further.

* 2022-08-026 **Colab Demo v0.2:** [<img src="https://colab.research.google.com/assets/colab-badge.svg" align="center">](https://colab.research.google.com/github/vijishmadhavan/Chehara-GAN/blob/master/%E0%A4%9A%E0%A5%87%E0%A4%B9%E0%A4%B0%E0%A4%BE_GAN.ipynb)

* 2022-08-026 [beta-v0.2](https://www.dropbox.com/s/iiqvfu58as8unz1/p6500.pkl?dl=1)

* 2022-08-023 [beta-v0.1](https://www.dropbox.com/s/d43p26ikzlxuyix/p4500.pkl?dl=1)

* 2022-08-022 Colab demo v0.1 [<img src="https://colab.research.google.com/assets/colab-badge.svg" align="center">](https://colab.research.google.com/github/vijishmadhavan/Chehara-GAN/blob/master/%E0%A4%9A%E0%A5%87%E0%A4%B9%E0%A4%B0%E0%A4%BE_GAN.ipynb)

* 2022-08-022 Fast.ai model training [url](https://github.com/aarcosg/fastai-course-v3-notes/blob/master/refactored_by_topics/CNN_L7_gan_feature-loss.md)

## Example Images

**64×64 -> 512×512**

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/0_41_inf-side.png)

**64×64 -> 512×512**

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/0_21_inf-side.png)

### Image samples

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/sidee.png)

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/1_7-side.jpg)

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/Adele_crop-side.png)

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/90050137-4da1-44cc-b64b-0b9efc813148-side.jpg)

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/ami-side.jpg)

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/0941881e-1b87-46d3-9b4e-10e9b8b4137b-side.jpg)

### Dataset

Dataset was generated using the below models.

-[GPEN](https://github.com/yangxy/GPEN)

-[GFPGAN](https://github.com/TencentARC/GFPGAN)

-[FFHQ](https://github.com/NVlabs/ffhq-dataset) can be used, but generating a dataset would help in dealing with real image problems.

-Generated dataset was cropped to facial features and then trained.

![Superstar](https://github.com/vijishmadhavan/Chehara-GAN/blob/master/compare/facial%20feature-side.jpg)

### Model

[beta-v0.1](https://www.dropbox.com/s/d43p26ikzlxuyix/p4500.pkl?dl=1)

[beta-v0.2](https://www.dropbox.com/s/iiqvfu58as8unz1/p6500.pkl?dl=1)

Use this [weights](https://www.dropbox.com/s/1818xtxblsyrv45/c2.pth?dl=1) to train it further.

- works well with color photos.

### Implementation Details
- Model architecture: [fastai v1 u-net](https://fastai1.fast.ai/vision.models.unet.html)
- Paired image2image training: [fastai v1 superres notebook](https://github.com/aarcosg/fastai-course-v3-notes/blob/master/refactored_by_topics/CNN_L7_gan_feature-loss.md)

### Limitation

- I call this a beta as it needs lot of improvement, But with minimum effort the results are good.

Chehara-GAN

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README