# Sprite DX - Pixel Alignment - Lab Notes 2

I trained a model to cleanup the pseudo pixel arts:

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1767902956846/d8f2d1b3-df7b-4df1-88bb-fdbec5fe230e.png align="center")

Left is 128×128 resampled generated images.

Middle is the cleaned up version on top of white bg.

Right is the cleaned up version on top of black bg.

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1767903097317/581ea2be-5605-4359-abd2-e53ee551426f.png align="center")

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1767903103568/b287b2a2-52c8-4c3a-8a56-1bc2bda87100.png align="center")

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1767903106794/75076a55-43b3-42af-8d13-22d3c8e63370.png align="center")

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1767903110709/92789d6e-70ba-4c57-898b-3b4577f0099d.png align="center")

From zoomed out view, I think these characters look decent. However, there are quite a bit of alpha errors.

1. We need more various data
    
2. See if we can use pretrained model