Posts: 664 +27
Forward-looking: New AI-based image generators like Dall-E have produced shockingly accurate and complex art based on fairly simple text commands. Some think they could change art production workflows forever. A new experiment demonstrates how AI image generation could work for game development.
An independent developer has released a short 2D shooting game they made using only AI-generated art. The game itself isn't very remarkable but shows some of the advantages and limitations of using AI to create art for a game.
Developer "Nao_u" built the title, called Shoon, over three days to see if Midjourney – an AI image generator similar to Dall-E – could create art for a game. Nao_u made a single level for a horizontal shoot-'em-up while generating the background, the player's ship, and the enemies with AI.
Even the title came from AI. When the developer asked Midjourney to create a shoot-'em-up logo, it generated letters appearing to spell "Shoon."
Nao_u built the ship models by feeding Midjourney text prompts related to Star Wars and Armored Core. The developer compared the process to playing gacha games because they had to pull Midjourney's figurative lever multiple times until it made something usable. However, Nao_u was impressed with how many variations the generator could make.
The background is a looped city-themed landscape with translucent edges and a perspective-adjusted ground texture. Nao_u wanted to make more levels with grassland, desert, and ocean backgrounds, but this would have taken longer.
Midjourneyで"成-た絵'使って横スクロールシューティングゲーム'作ってみた pic.twitter.com/M6HUMhzKkW— Nao_u (@Nao_u_) August 13, 2022
One of the environment art's biggest limitations is that it's a single, static image. Nao_u couldn't separate elements like the clouds, buildings, and vegetation into the foreground, middle ground, and background to create the sense of perspective common in 2D games.
That drawback forced Shoon to be a horizontal shooter instead of a vertical one, which would require a top-down view of the environment. AI generators aren't known to make images with that perspective, and it isn't clear how well they can.
Another limitation of static-AI images is that the character sprites can't animate. A shoot-'em-up only featuring static space ships can work around that problem, but Nao_u also created bug-like enemies by obscuring their legs, which wouldn't move.
The lack of animation makes games featuring human characters much harder to build with AI-generated art. However, Nao_u notes emerging advancements in 3D model AI-generation allowing for animation, expanding the possibilities for game development.
Building things like textures and models can be one of the most expensive aspects of game development. AI generation probably won't replace all of it, but as it progresses, it could make certain parts easier.