Does not support batching in the differential img2img version
Thank you for the excellent tutorials. It seems like the ip_adapter patch of the tokens are really broken and its not allowing any easy fix to simplify the batching process for the differential_img2_img pipeline. Aret here any fixes?
Hi, thank you for your kind words. I don't really use batching while generating so I never noticed it. Currently I'm not happy with the way IP Adapters are implemented since it's really complex and hard to follow when you use multiple IP Adapters with multiple images, masks and scales, if you add batching over it, it's not strange to me that it breaks.
Probably I'll take a look into it for modular diffusers which is our next feature release, but we don't have a timeframe for it yet. I will also update all the guides (which are now kind of old) and make new ones that uses more the power of the modular part.
I just need a method to batch outpainting with quality at SDXL or better. There used to be a method that Runway used which I think they have now removed we only have the v1.5 version out now. My only goal is to simply outpaint left and right of the input 256px or some arbitrary number like you do in the differential img2img pipeline. Are there any existing ones with relatively high quality which currently already exist?
with the other guide I did, the image fill one, at the end I comment that it also works for outpainting which in my own biased opinion is the best one there is, I haven't seen any other that has better quality. But if you want, you can also use the Flux Fill model, it also has the capability to do outpainting.
Both of those are better than the old guide with differential diffusion I did.
There is a user that used my guide and extended the space here to do outpainting if you want to test it.
Also there is one that does the same with Flux Fill here