README.md

---
license: openrail
datasets:
- balgot/stylegan3-annotated
language:
- en
metrics:
- mse
tags:
- face-generation
- stylegan3
library_name: pytorch
---

# Text-to-StyleGAN3 Latent Space Translation

This model was created as a part of the project for FI:PA228 (Masaryk University), 
inspired by this paper: [Face Generation from Textual Features using Conditionally trained Inputs to Generative Adversarial Networks](https://arxiv.org/abs/2301.09123)

It was trained on the generated dataset from BLIP and StyleGAN3. See the [corresponding notebook](https://colab.research.google.com/drive/14rDcCc0Xr1L1Ax3aKezEhmcn81vXGVQ7?usp=sharing)
for further details.


## How to use:


```python
import torch.nn as nn


# for now, the model class needs to be defined, so...
class LaTran(nn.Module):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.pipe = nn.Sequential(
            nn.Linear(384, 512),
            nn.ReLU(),
            nn.Linear(512, 512)
        )

    def forward(self, v):
        return self.pipe(v.unsqueeze(1))



# Instantiate and load the model
dev = ...  # device to use
PATH = "translation_model-sd.pt"  # local path
model = LaTran().to(dev)
model.load_state_dict(torch.load(TRANSLATION_MODEL, map_location=dev))
```

## Demo

For the demo of the whole pipeline, or how this model helps to generate a final image,
visits [text-to-stylegan HF space](https://huggingface.co/spaces/balgot/text-to-stylegan3).

## Examples

* Prompt: `attractive young woman, blond hair`
  ![image of attractive young women](attractive_young_woman_blonde.png)


  

* Prompt initial: `cute young boy, blond hair, blue eyes, smiling`
* Prompt second: `old man, short gray hair, glasses, wearing hat`
  <img src="https://huggingface.co/balgot/bert-2-stylegan3/resolve/main/young2old.gif" width="200" height="200" />