Nice to see a ComfyUI video that starts from the basics and not some overly complicated workflow with 15 custom nodes jusrt to upscale. Clear and to the point !
With the upscale with model method, how do you adjust its denoising strength and steps like you could with the latent upscale? Because in a1111 webui you can adjust these parameters not only with the latent upscale but with the model upscale too. Thanks.
why not just start out with a 1024 latent image? i can see wanting to upscale an image that already exists but what's the point of creating small only to immediately enlarge it? maybe i'm missing a practical usage for this workflow. either way i like your straight to the point style. thanks for the videos.
1024 x 1024 latents in SD 1.5 tend to encourage the sampler to duplicate elements of the prompt. So, if you ask for "a chair", there's a pretty good chance with many seeds that it'll give you TWO chairs.
In regards to "Load Upscale Model": what all do I need to know, i.e. do I need to go to civitai/huggingface and download upscale models? Where do I put them, in the checkpoints folder?
Hey GES1985 - you can place upscale models in the ComfyUI\models\upscale_models directory. Models are available from multiple places as you mentioned - civitai, huggingface, openmodeldb.info.
I am used to drawing dozens of small images first, then selecting the satisfactory ones, using SD's PNG INFO to recognize the information of the small images, and then using Hire. fix to enlarge them; How to operate in ComfyUI?
I have a question, does either of these methods add detail/quality if you just want an image of the same size? Like, I don’t want a giant image, I just want it to look crisp and without blurry eyes or other skewed fine details. I’m using SDXL so the sizes it generates are fine for me size wise but I see automatic1111 users say high res fix can fix eyes and small details like that.
You can actually do a downscale but it still “upscale” in that the latent will be bigger, add detail and then go back to pixel space (what you actually see) in the original resolution. Look into kohya’s packages for stuff like that.
In automatic just do hires fix or img2img and then set the ultimate sd upscale to 1 and it will do the latent denoising and spit back out a more detailed image.
You shouldn’t do that. It will either crop the image and be unrecognizable or it will stretch and look goofy. You have to upscale to something in the same ratio.
Unfortunately this is nowhere near real hires-fix but it'll do for now. Thanks for a great explanation UPD: ComfyUI docs have a workflow for hires fix with a couple extra steps
Nice to see a ComfyUI video that starts from the basics and not some overly complicated workflow with 15 custom nodes jusrt to upscale. Clear and to the point !
Thank you for the kind words.
Couldn't agree more, this is so useful information and structure! Subscribed.
YEESSS Finally!
Finally a guide that goes from the basic to the specific point. Subscribed!
Thank you!
With the upscale with model method, how do you adjust its denoising strength and steps like you could with the latent upscale? Because in a1111 webui you can adjust these parameters not only with the latent upscale but with the model upscale too. Thanks.
There’s a custom node you can install to do that. I think ultimate sd upscale has the ability to do that.
Also I wouldn’t do upscale by model it’s not as good as using the latent through the ksampler
Great and well laid out video again, and filled with info! Thanks!
Thanks so much again Robert!
That was so awesome. It took me a couple tries finetuning the settings since I'm used to SD 1.5 but this was exactly what I wanted.
why not just start out with a 1024 latent image? i can see wanting to upscale an image that already exists but what's the point of creating small only to immediately enlarge it? maybe i'm missing a practical usage for this workflow. either way i like your straight to the point style. thanks for the videos.
1024 x 1024 latents in SD 1.5 tend to encourage the sampler to duplicate elements of the prompt. So, if you ask for "a chair", there's a pretty good chance with many seeds that it'll give you TWO chairs.
In regards to "Load Upscale Model": what all do I need to know, i.e. do I need to go to civitai/huggingface and download upscale models? Where do I put them, in the checkpoints folder?
Hey GES1985 - you can place upscale models in the ComfyUI\models\upscale_models directory. Models are available from multiple places as you mentioned - civitai, huggingface, openmodeldb.info.
Also usable with mp4 as input? Have some memory issues all the time with upscale mp4's...
Under-rated channel. Ty!
I am used to drawing dozens of small images first, then selecting the satisfactory ones, using SD's PNG INFO to recognize the information of the small images, and then using Hire. fix to enlarge them; How to operate in ComfyUI?
I have a question, does either of these methods add detail/quality if you just want an image of the same size? Like, I don’t want a giant image, I just want it to look crisp and without blurry eyes or other skewed fine details. I’m using SDXL so the sizes it generates are fine for me size wise but I see automatic1111 users say high res fix can fix eyes and small details like that.
You can actually do a downscale but it still “upscale” in that the latent will be bigger, add detail and then go back to pixel space (what you actually see) in the original resolution. Look into kohya’s packages for stuff like that.
In automatic just do hires fix or img2img and then set the ultimate sd upscale to 1 and it will do the latent denoising and spit back out a more detailed image.
What if u have 832x1216 image (SDXL 1.0)? How to upscale it to 1080x2800? getting kinda bad quality
You shouldn’t do that. It will either crop the image and be unrecognizable or it will stretch and look goofy. You have to upscale to something in the same ratio.
Unfortunately this is nowhere near real hires-fix but it'll do for now. Thanks for a great explanation
UPD: ComfyUI docs have a workflow for hires fix with a couple extra steps
thank you for sharing - will have to take a look at the workflow
당신의 강의영상이 많은 도움이 됐습니다.
대단히 감사합니다....
nicely explained, clean workflow.. keep going 👍
Thank you!
How I can setup hires_steps param when upscaling? This is useless video for me..
great video G
Appreciate x2!
Why do they call it a hi-res fix? It's not fixing anything, it's just a normal upscaling.
Very helpful, thanks.
this isnt hires fix, this is upscaling. hires fix is using a adetailer, face detailer etc
great video
Thanks
ComfyUI is stupidly complicated
Nodes often become more advantageous as one becomes accustomed to utilizing them.
At first it sucks but when I actually wanna try and get good images I use comfy
great video G
Appreciate it