Hi, I’ve finished drawing my autotiles for RPG Maker (48x48). They’re just the floors and walls, but technically, it's the most tedious and least enjoyable part of the work, so I’m happy to have it done for now.

Next, I should move on to structures and furniture. Once I’ve finished this process, I’ll be able to map out an environment/room and take it to training. I’ll talk a bit about this process because it might be interesting. Attempting to generate rooms and backgrounds for a scene is impossible without connecting this to a workflow. My tiles have a low-saturation tone and very cold colors, and capturing this with generative AI is impossible unless… unless we do the following. I want to make it clear that this is an initial test, and I’ve used test elements to do it:

This image is a sample from a training process where I separated all the elements in the image and rearranged them in different positions and sizes. The dataset for this training consisted of 12 similar images, using the same background and the same characters.

Results:

As you can see, this training not only manages to recreate the characters with at least 90% accuracy but also successfully captures the background colors.

This test opens the door to connecting my tiles with generative AI. My process will involve creating scenes, training generative AI with them, capturing the tones and colors of my tileset without any characters involved, then generating backgrounds and trying to recreate them with the tileset. Once I have these backgrounds, I’ll retrain the model with backgrounds that fully match the colors and environments I’m creating in pixel art.

It’s a bit exhausting, but I’m trying to clearly document how I’m converging manual work with generative AI.

Soon, once I finish A1, A2, A3, and A4 in RPG Maker (floors and autotiles), I’ll share them, along with the character model and their sprites. For now, this is a lot of progress I’ve made over the weekend, and it’s quite motivating. See you soon for now!