MaraScott / ComfyUI_MaraScott_Nodes

A set of community node including a universal bus and a large Upscaler/Refiner
https://www.marascott.ai/
Other
113 stars 13 forks source link

does not support tiled controlnet when upscale image #47

Closed frankyifei closed 4 months ago

frankyifei commented 5 months ago

I will apply controlnet with original image to each tile, and end up with something like this. It does not work even if you set the output size as same size. ComfyUI_temp_ypfak_00001_

MaraScott commented 5 months ago

original issue https://github.com/davask/ComfyUI_MaraScott_Nodes/issues/45

MaraScott commented 5 months ago

Do you know any Upscaler which allow controlnet conditioning as input ?

Could you use the below workflow, to finetune the image using the controlnet, in order to test the McBoaty against the expected ControlNet Image ?

ComfyUI_temp_pfjou_00006_

workflow-controlnet

frankyifei commented 5 months ago

Ok, ultimate sd upscaler support controlnet and one of the reason to use tile controlnet is to add detail while not change image.

MaraScott commented 5 months ago

Thx, I ll check ultimate SD upscaler then :)

MaraScott commented 4 months ago

@frankyifei , after debate on fern discord, we come to the point that upscaling comes far down after the control net and using an upscaler/refiner like McBoaty as the goal to refine it while increasing the size (or not) so it won t involved control net.

I will give you @treeshark aka Rob Adams answer on that topic

I doubt if any tiled upscale would. I could make a flow that worked with control net, you would have to divide up the canny or depth for each tile. So tile one would get the appropriate bit of guidance. I don't see the point though, refining and upscaling is for after you have resolved your image. Control nets and Adapters are best applied at a lower res.

So even if I do want to make McBoaty compatible with whatever comes in, for now - because of a lack of free time - I won t include it.

I hope you will still enjoy what McBoaty can bring to you up to .35 / .45 denoise you should enjoy it perfectly and for sure you need to provide an image without controlNet in it.