AI Materials, the Future?

[WAS]

Currently AI is in it's infancy, though with Stability AI, things have taken a very large leap. That being said, it's only a matter of time until Stability AI and others can do perfect cohesive textures. It's already not far off. And with the filters I've added to my version of SD (https://rebrand.ly/easy-diffusion), can already output seamless textures with depth maps which can be doctored for height maps.

The prompt wasn't even engineered here, just simple "Desert rock seamless texture" (which of course  wouldn't be seamless without my tiler, but tells the AI more or less what I want)

[René]

That's clever. I guess the next step will be to be able to generate 3D models based on prompts. That will be more difficult I guess, as 3D is less forgiving due to the high degree of accuracy required. Actually, I don't have a good understanding of how text to image might work. I imagine some sort of intelligent collage is created, similar to how matte painters incorporate photos seamlessly into their work

[WAS]

That's clever. I guess the next step will be to be able to generate 3D models based on prompts. That will be more difficult I guess, as 3D is less forgiving due to the high degree of accuracy required. Actually, I don't have a good understanding of how text to image might work. I imagine some sort of intelligent collage is created, similar to how matte painters incorporate photos seamlessly into their work

There is already work in that area, using the AI knowledge to interpret 3D meshes. There is an example that uses I belive voxel points to create a 3D world from an image that can be navigated, and stuff like this: https://towardsdatascience.com/using-ai-to-generate-3d-models-2634398c0799

Once a principle is fine-tuned to something that works from reference, they will likely put together a massive model that you can try to query any model from like: "A teddy bear riding a toy train"

The amount of training that went into Stable Diffusions model is mind-boggling. Altogether the training that took place would have taken a single person 62 years to complete on their own (and I think this was calculated for a current gen consumer 3090). o.o

[Dune]

That is quite something (to look forward to)!