Hi, I’ve discovered this channel a while ago and it’s great to see that someone does AI benchmarks with AMD GPUs!
I’ve had this idea for a while to make an improved set of tiles for the free and open source game OpenTTD. It’s an isometric 2D game where players grow cities and manage their own transport company. The game supports mods, including new tile sets. Currently there are two popular sets of tiles that players can download from the game’s menu:
- OpenGFX - this is what the game comes with by default (GitHub - OpenTTD/OpenGFX: 8bpp Graphics Base Set for OpenTTD)
- Zbase - a cleaner looking set of tiles that was rendered in Blender
(GitHub - gravitystorm/zbase: A custom version of the zbase openttd graphics pack - not the original repo it seems, but the only one I could find)
My idea was to generate a new, more modern looking set of tiles entirely using AI. In theory it should be possible to generate, for example a picture of a train with some image model (example prompt: “a realistic 3D render of a steam train locomotive”), then feed that picture to an image editing model, like Qwen Image edit, to generate consistent views of this object from different angles.
The first model can be anything, Qwen Image, Z-Image, but Qwen Image Edit can be used for that first step too, which is pretty cool.
Another possible approach is to use an AI model that can generate a 3D mesh from an image, then render that in Blender from multiple angles. This is what the Zbase project did, except they didn’t use AI to generate meshes and textures (it was done over a decade ago). They had a Blender scene that would render every model from multiple angles on a plain background, with the same lighting and those pictures could then be used in game. So I could use (for example) Qwen Image to generate a picture of a train, then feed that into something like Hunyuan 3D 2.1 to turn it into a 3D model. Then somehow setup a Blender scene to render it all for me from all the right angles.
My findings so far
So far I’ve been unable to get consistent angles with Qwen Image Edit 2511, even with one of the new multi angles loras (fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA · Hugging Face). So that’s not good. So right now I’m experimenting with feeding the existing Zbase images as reference to Qwen Image Edit. I convert them into depth maps and Canny edges, I’m not sure if that will help. It’s possible that maybe a custom lora is needed for this task.
What about the 3D approach? The meshes generated by Hunyuan 3D 2.1 seem to need a lot of cleanup to be usable and it doesn’t seem to generate textures. This is my first time using those this type of AI model, though. This still seems doable, but it would be more work than I was hoping and the image generation approach seems more fun to me (if it can work). There is also the recently released Trellis 2 model, maybe it would be better at this (if it works on AMD cards).
Other possible projects
There are some cool 3D free and open source games like Super Tux Kart and 0 AD. Textures for 3D games are a more complicated thing. One texture requires multiple images - normalmaps, etc. So I’m not sure if AI can really be used for that, at least with the free and open source tools that we currently have available. But maybe I could do something simpler like upscaling existing textures in those games with an upscale model like SeedVR2 to 4k resolution to make them look sharper. I don’t know if those two games support those kind of mods (I kinda doubt it), but since they are free and open source, I could change the textures, compile a new build and distribute it.