Generation of game assets in the spirit of Red Alert using Stable Diffusion

Emmanuel de Maistre, CEO and co-founder of the Scenario startup, demonstrated what level of gaming assets can be achieved from neural networks today.

At the same time, on his Twitter page, he shared how he achieved the desired result by creating a bunker in the style of his favorite game – Red Alert.

At first, Emmanuel generated more than 200 buildings in the spirit of the game using Midjourney.

Of these, he chose 16 images that were quite different from each other, but at the same time remained within a certain consistent style.

They have become references for training Scenario_gg, built on a Stable Diffusion model focused on generating game assets (the model itself, by the way, is not yet available, the release of the alpha version is promised in the coming weeks).

After the model passed the training, Emmanuel began to ask her commands to create various types of buildings based on the received references by text query.

According to Emmanuel, the neural network successfully coped with most of his commands. For example, this turned out to be a nuclear power plant:

The neural network offered such variants of sprites at the request of the “factory”:

And here’s an example — a refinery:

But Emmanuel was dissatisfied with the Soviet bunker.

Therefore, he decided to generate a set of bins based on one specific model. That is, the text query “bunker” was saved, but at the same time the neural network was offered to focus only on one specific image.

Then Emmanuel changed the original image, leaving the previous request. Bunkers have changed a lot. They no longer looked like a factory building, but they also resembled bunkers only remotely.

In the course of further experiments, the co-founder of Scenario achieved satisfactory results.

Then, having already taken the model that suited him as a reference, he began to customize it with the help of clarifying queries. This will result in a bunker for the query “isometric bunker, realistic, soviet flag, red, video game”:

And such by request “isometric bunker, realistic, USA flag, blue, allied, video game”:

By changing the original image and the text query, he achieved the generation of a set of bins in different environments:

There are also other examples in the original thread. For example, what options does the neural network offer if the developer wants to generate a destroyed radar in the style of the Fallout game and not only.

This is not the first Emmanuel thread dedicated to the capabilities of neural networks in terms of generating game assets. At the beginning of November, he was creating a bottle with a potion.