[Feature Request] Clip comparisons

[EtherealIntellect]

Okay, so i only tried this in a really hacky way so not sure if it's valid, but:

An option to check the clip embedding of two images, subtract them before converting to text pompt styles to get both a positive and negative style prompt that might help with turning image1->image2. (Maybe even hint at how many (())[[]]?)

I've seen great success with how good the embedded clip search works on lexica, so feel like more useful things like this could be done before converting it back to text

As a bonus maybe even batches of images could be averaged and compared for styles

[ryukra]

I researched this topic too, I actually think clip end2end would be the way to do this.. not generating a prompt(because you lose a lot of information) and instead use the embedding itself to generate similar pictures. But im not 100% sure how to do it yet.

[EtherealIntellect]

There's https://github.com/justinpinkney/stable-diffusion which might be enough to avoid text prompts all together, but this one would also need the seed noise reversal/reconstruction improvements lately (unless image1 is stable diffusion generated and the seeds between this and that model are close enough, that one doesn't even have a seed input by default)

Get clip image1, get noise seed, get clip image2 and run it on the same seed