Training a Neural Network on the «Ceramic Tile» Style

The main idea and goal of the project was to train the Stable Diffusion neural network to create images in the style of decorative ceramic tiles, replicating the specific glossy effect (visible in the images presented below). A total of 36 images were collected. Because I was very impressed with how the insects were depicted, I decided to dedicate the entire series to them.

I strictly followed the instructions in the collab to prevent any errors.

I compiled all the images into a Google Drive folder to make it easier to use in the code, and then I imported them.

Next, following the instructions, I checked everything to make sure it was working and had uploaded properly.

Creating a suitable name for my style

After a few more lines of loading libraries and other necessary elements, I move on to the core learning module. This section of code took me about 30 minutes to process.

Maintaining the style from the Hugging Face website, where we also sourced the token.

Done! Now you can write prompts, which I have done.

Note: I did not describe every block because I am not the code creator and tried to follow the instructions; instead, I highlighted the most crucial stages in my opinion.

I think it turned out great in capturing the texture of the original tiles because that gleam is present. The only thing is that some of the images make them slightly more realistic than they should be on the tile, but I think this is because there were only 36 training images.

The other decorative elements also turned out very well—almost symmetrical, although you can see artifacts. Overall, I think this learning experience was successful; a prompt consisting of only three words can create an image nearly identical to the originals.