I'm actually curious about this one in particular, because it seems to vaguely match either overall shape and color or the type of object something is. I kind of feel like it must be doing some layered thing to composite the images and is only generating parts of it dynamically because of specifically how it is both good and bad at mimicking images.
No, there are publicly available models that are about this good at composition with just text prompts. Having an image prompt makes things significantly easier. Their model might be doing some depth masking, but it doesn't look like it from the examples I've seen.
They just have to have images they’re most closely matching these things up to right? I’m impressed with the lack of body horror for AI art
I'm actually curious about this one in particular, because it seems to vaguely match either overall shape and color or the type of object something is. I kind of feel like it must be doing some layered thing to composite the images and is only generating parts of it dynamically because of specifically how it is both good and bad at mimicking images.
No, there are publicly available models that are about this good at composition with just text prompts. Having an image prompt makes things significantly easier. Their model might be doing some depth masking, but it doesn't look like it from the examples I've seen.