comfyanonymous
[Feature Request] Is there any plan to implement reference_only（sd webui controlnet） nodes?
Compared with the experimental node and sd webui controlnet, the effect is not very good.
IMO, I don't think there will be any further development for "reference_only." Since there are already more advanced tools like IPAdapter, the merits of "reference_only" seem to be diminished.
IMO，我认为“reference_only”不会有任何进一步的发展。由于已经有像 IPAdapter 这样更先进的工具，因此“reference_only”的优点在于：似乎有所减少。
But I use the same conditional control, and it seems that the effect of reference_only is better than that of ipadapter. Is there something wrong with my use of ipadapter? I don’t know whether there is a lot of development to implement reference_only, but it seems that ipadapter cannot completely replace reference_only.
Your scheduler isn't set to karras like it is in the below image for one thing, which matters a lot, and your empty latent doesn't match the input image so it's making up the edges the image doesn't cover which will bleed into the input.
Your scheduler isn't set to karras like it is in the below image for one thing, which matters a lot, and your empty latent doesn't match the input image so it's making up the edges the image doesn't cover which will bleed into the input.
thank you for your advice，But I retested it, and the effect is still not very good. The main reason is that it cannot maintain style consistency, even though its character characteristics are
It seems unlikely much more will be done: https://github.com/comfyanonymous/ComfyUI/issues/661 If you look at the source you'll see why, it's all mixed up and poorly named within a general purpose hooking module: https://github.com/Mikubill/sd-webui-controlnet/blob/01e4574d8e01fa00628fdae0c8215283a1c36a8d/scripts/hook.py
I messed with IPAdapter for a while since I'd never used it before and had better luck plugging it into both the apply IPAdapter and a CLIPVisionEncode -> unCLIPConditioning. I tried on a detailed and fairly complex scene and got the best results with a detail LoRA tacked into the chain, you probably won't need that for anime (mine was tending towards looking too cartoony without it). With both of those on and set to weight 1 you shouldn't get a ton of variation from the base picture unless you add a prompt, too.