HOW TO USE ARTICLE HERE
Might not work with newest comfyui with subgraph features.
Update: 06/10
There have been several requests on how to make controlnet work with the workflow, here is example for it. I have somewhat old comfyui, if some nodes do not work, you might need to remake them.
I recommend using same size image for the controlnet, as you use for generation. Then you can copy paste them to the load image for masks for easy masking.
The size for generation is taken from first mask!
On this example controlnet is only used for first sampler, if you need it for some reason for the second sampler too, you can just wire it from the controlnet.
Update: 01/04
Added version that does not use CGEM nodes, as people seem to have problems getting it working. Only did fast testing with the non cgem version and seemed to work. Let me know if you notice some problems in it, i am using slightly old version of comfyui.
Simple example workflow for regional prompting in comfyui.
tested to work with XL, might work with other models too.
Description
Initial version
FAQ
Comments (10)
This works pretty good. Easy to use, runs fast.
My only real issue is that as far as I could tell, there's no way to assign specific loras to specific masks. This would help a lot for avoiding contamination between characters. sometimes just isolating trigger words isn't enough, some character loras will overpower everything just by being present.
Thanks, i usually try to make stuff i release easy to use. (～￣▽￣)～
Sadly what comes into loras, i can't think of fast method/nodes to implement regioning of whole lora. With the current nodes in the workflow, i don't think it is possible outside the triggers on the masks. I wonder if some node that calls the lora with texts would work..
@PotatCat Likke0703's "Regional Conditioning with AttentionCouple(LoRA supported)" workflow does it, but yeah it is very slow, like over half an hour, but that's likely because I have only 8GB VRAM.
@DizBabes Checked it fast, seems to use conditioning nodes. Those slow down generation more you add them, but should not slow down that much. Are you using --lowvram arg?
@PotatCat no i am not. i tried with and without it, as well as with and without the --force-fp16 arg, but it was always very slow.
@DizBabes Took another fast, i see it using some controlnets and sam model for something. You could try disabling those and see if it is still slow. Those are at the top. Usually it loads those to vram.
Also at what point does it get slow? What is it doing in the CMD. If you open task manager -> go to performance -> at the bottom there is your GPU. Does it say the GPU memory is full and there is something in the shared GPU memory?
If the generation gets into shared memory that is where it will slow down. Also if you run out of normal RAM it will go into page file, if that is in slow HDD it will be even slower. (～￣▽￣)～