justinpinkney / stable-diffusion

MIT License
1.45k stars 266 forks source link

Novel feature for finding the noise - maybe can be useful for style transfer? #7

Closed EtherealIntellect closed 1 year ago

EtherealIntellect commented 2 years ago

https://www.reddit.com/r/StableDiffusion/comments/xboy90/a_better_way_of_doing_img2img_by_finding_the/

Just saw this and been thinking, if you take clip1 embedding from image1, reconstruct the noise1, take image 2, find clip2, and recreate from noise1 to get a style2 result. Still only just read about it so haven't thought it through, but the reconstruction idea seemed very useful.