Still not possible to make a sharp, accurate, high-resolution latent upscaling to say 4x. I think the technology is not there yet. Maybe control net will eventually help us in that direction. The fractural results you may get in latent upscaling with high denoise values can't be avoided at the moment. We need some sort of regional upscaling that understands the whole image and its components and adds reasonable details to each component separately rather than trying to add details in one go for the whole image, like upscaling layers. Like it upscale the background, then go for the body, then the clothes, then the face, then face fetures, then eyes, then iris, then pupil... and this is done to the lips , to the nose, to the nipples, to the finger nailes, to the skin, to every small tiny component of the whole image. Like a branching upscaling.
What it really needs is tiled conditioning that generates its own map of the image and then passes it along to upscalers in order to isolate conditioning to the correct locations on the image. this would stop it trying to generate the subject of your image in a square in the corner when you tiled it. I had some success using 'blip' to detect the content of each tile before upscaling. but blip is very limited.
That's a good idea. I'm also using this method, and my images usually reach 8K, 8000x6000 resolution, without losing sharpness even at 500% zoom. So, instead of combining it with "upscale image," can I replace it with "upscale image by" to easily choose the scaling factor, rather than having to calculate how much to upscale by, and then connect the "upscale image by" node to the 4x-ultrasharp node?
whatever works for you really, I don't like upscale image by for one reason. I can't do math so it becomes hard to know what pixel sizes ill get out the other end. all the numbers get jumbled in my head by dyslexia, managed to fix it with writing but didn't know I had it for math until it was too late to correct :P Just go for it!, don't even bother to ask, just try it and see if it works, im not some SD god I make mistakes XD
Excellent, thanks. Can you please do a XY plot with SDXL ? I often do LoRA training epochs on X axis and a few seeds on the Y axis to establish training success. Been doing this A1111 for a while but I really would like to port everything over to comfy and I just can't make this work.
@@ferniclestix that would be amazing and useful to many, doing cfg variations, step variations (especially base/refiner ratio) etc. i don’t believe that would be beyond you at all!
had a look, I mean... I plugged everything in correctly and it outputs nothing useful lol.... gonna be waiting a while if you want a tutorial on that, it may be bugging out but followed workflow on the actual github and it doesn't output what it claims to.... which doesn't make sense to me. *shrugs*
Very usefull. Great tip with scrolling and not losing the text in the nodes too.
Still not possible to make a sharp, accurate, high-resolution latent upscaling to say 4x. I think the technology is not there yet. Maybe control net will eventually help us in that direction.
The fractural results you may get in latent upscaling with high denoise values can't be avoided at the moment. We need some sort of regional upscaling that understands the whole image and its components and adds reasonable details to each component separately rather than trying to add details in one go for the whole image, like upscaling layers. Like it upscale the background, then go for the body, then the clothes, then the face, then face fetures, then eyes, then iris, then pupil... and this is done to the lips , to the nose, to the nipples, to the finger nailes, to the skin, to every small tiny component of the whole image. Like a branching upscaling.
What it really needs is tiled conditioning that generates its own map of the image and then passes it along to upscalers in order to isolate conditioning to the correct locations on the image.
this would stop it trying to generate the subject of your image in a square in the corner when you tiled it.
I had some success using 'blip' to detect the content of each tile before upscaling. but blip is very limited.
i love your videos. great stuff.
Great keep em coming Fern, Can you do a series on Animate Diff ?
sooooon, first modular layout tutorial then a couple of modules, then animate diff
Great 🙂@@ferniclestix
When displaying the menus you should zoom in more, because here it is unreadable (or very close to it).
yeah, working on that more in the newer tutorials, always learning :D
That's a good idea. I'm also using this method, and my images usually reach 8K, 8000x6000 resolution, without losing sharpness even at 500% zoom. So, instead of combining it with "upscale image," can I replace it with "upscale image by" to easily choose the scaling factor, rather than having to calculate how much to upscale by, and then connect the "upscale image by" node to the 4x-ultrasharp node?
whatever works for you really, I don't like upscale image by for one reason. I can't do math so it becomes hard to know what pixel sizes ill get out the other end. all the numbers get jumbled in my head by dyslexia, managed to fix it with writing but didn't know I had it for math until it was too late to correct :P
Just go for it!, don't even bother to ask, just try it and see if it works, im not some SD god I make mistakes XD
How to apply lora with SDXL?
will be next tutorial :P working on it right now actually.
done :P th-cam.com/video/hJXtW44QBcs/w-d-xo.html
Excellent, thanks. Can you please do a XY plot with SDXL ? I often do LoRA training epochs on X axis and a few seeds on the Y axis to establish training success. Been doing this A1111 for a while but I really would like to port everything over to comfy and I just can't make this work.
Ill look into it, ive never really used the XY plot stuff in comfyui, It may be beyond me
@@ferniclestix that would be amazing and useful to many, doing cfg variations, step variations (especially base/refiner ratio) etc. i don’t believe that would be beyond you at all!
had a look, I mean... I plugged everything in correctly and it outputs nothing useful lol.... gonna be waiting a while if you want a tutorial on that, it may be bugging out but followed workflow on the actual github and it doesn't output what it claims to.... which doesn't make sense to me. *shrugs*
@@ferniclestix this is the sort of thing I mean th-cam.com/video/tRsFmNI-yj0/w-d-xo.html
Yeah, was trying out this XY node, couldn't get it to do the thing :< github.com/LucianoCirino/efficiency-nodes-comfyui