TLDR; 'object made from texture'
I spent a bit of time training PixelWave on textures. I collected hundreds of pictures of textures, which I used to train LoRAs. The image may have been something basic, like a cube or sphere with the texture applied. I trained some LoRAs with 1 image.
'cube made from water surface'
Then I used those LoRAs to generate images of other objects made of that texture. You get a lot of images that don't work, but every now and then it works, and you have something that is made out of that texture. This then goes into the next batch of training images for the next LoRA.
'man made from frosted glass'
I had this idea, if I could train enough textures into the model, will it learn how to make things out of textures that were not trained?
Well... SDXL proved to be pretty stubborn and turns out it can't just imagine something it hasn't seen before. But I persisted and kept training. The hardest textures to train were things like water, fire, smoke, things that aren't solid. Over 300 textures have been added to the model. I couldn't add the LoRA's into the model too strong or it would start to cause issues like duplicate limbs or a weird wet look being applied to everything. I think the model now has some understanding of how to apply textures that weren't trained, maybe through sheer amount of 'object made from texture' prompts it has been trained on.
`man made from finger print smudge'
It might take a few tries to get the result you want, but try it out for yourself and share your creations!
'man made from leaf veins'
'car made from spider web'
'car made from grilled meat surface'
'car made from cinnamon sticks'
'woman made from iridescent oil slick'
'pig made from spiky sea urchin'