# Training uv texture maps with Flux

Author: Raphael Vorias
Published: 2024-12-01
Source: https://vorias.digital/posts/2024-12-flux-uv-loras

I found some cool cigarette packs in the wild [🗗](https://opensea.io/collection/the-vintage-cigarettes) and wanted to train a Flux LoRA on them so that I could generate new designs.
The problem is they are all in different view arrangements.

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792006/2024-12-flux-uv-loras/pack_hollywood_green_tvbfdl.png)
![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792008/2024-12-flux-uv-loras/pack_hollywood_red_dk1k6r.png)

So I had an idea: first project them all to the same uv layout and **then** train a LoRA on them.
Would this work? Would Flux maintain the uv layout and know where to put text and info?

# Step 1: projecting all samples to the same uv layout.

In blender, I created the following model:

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792002/2024-12-flux-uv-loras/blender_1_qvl6ed.png)

The red edges indicate uv seams. They tell blender where it can *cut* the texture when unwrapping it.
The unwrapped uv map looks like this:

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792002/2024-12-flux-uv-loras/blender_2_h7d5xb.png)


I named the target uv map "SimpleBake", because I will be using the SimpleBake plugin and it will automatically search for uv maps named like this.
Then, I created a second uv map and activated it by pressing the render icon. 

The second map will contain a specific uv map that is different for each sample.

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792003/2024-12-flux-uv-loras/blender_3_gdw4bq.png)

Before you see anything, you also need to set up the material using the specific uv map:

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792003/2024-12-flux-uv-loras/blender_4_ympezm.png)

Using the second uv map and the material, we can now load a sample photo and we can start to drag the uv faces on the left to the right spots.
You have to eyeball it a bit and keep checking the result on the right all faces are mapped correctly.

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792010/2024-12-flux-uv-loras/blender_5_tqphfw.png)

And now the trick: we will bake the second uv map to the first uv map called SimpleBake. As mentioned before, I'll be using the SimpleBake plugin. 

The settings are: 

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792003/2024-12-flux-uv-loras/blender_6_ncnqwi.png)
![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792004/2024-12-flux-uv-loras/blender_7_kyes9p.png)
![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792004/2024-12-flux-uv-loras/blender_8_rdjkbl.png)

The cool thing is that this will be super fast since SimpleBake can use the Eevee renderer to bake objects.

And voila! I have now a transformed uv map:

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792006/2024-12-flux-uv-loras/blender_9_dqesia.png)

Rinse and repeat this for all training samples and kickstart a Flux LoRA.

# Step 2: Training a Flux LoRA

I used [Ostris/ai-toolkit](https://github.com/ostris/ai-toolkit) to train the flux LoRA, using pretty default settings.
Here are some results:

![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792007/2024-12-flux-uv-loras/pack_realms_uwyu6u.png)
![](https://res.cloudinary.com/dw8lz4sw6/image/upload/v1736792007/2024-12-flux-uv-loras/pack_brother_kldgjq.png)

# Step 3: Putting them on a new model

This is a quick python endpoint to swap textures and then serve the glb file. This is the only way I could get the textures to work on a new model in a live browser runtime.
(I needed this to display on the fly in a [glif.app](http://glif.app) artifact.)

```python
from pathlib import Path
from gltflib import GLTF
import shutil

DONOR_PATH = Path("./data/cigs.gltf")

def get_glb_with_texture(new_texture: Path):
    gltf = GLTF.load(DONOR_PATH)  # load method auto detects based on extension
    
    shutil.copy(new_texture, "./data/pack_001_Bake1_PBR_Diffuse.jpg")
    return gltf

def save_glb(gltf: GLTF, path: Path):
    gltf.export(path)

if __name__ == "__main__":
    glb = get_glb_with_texture(Path("tests/GdJXYGJWcAA2yR_.jpeg"))
    save_glb(glb, Path("tests/cigs_with_texture.glb"))
```