It's extremely amazing how your simple workflow has given me the results that others with hundreds of nodes and models and a huge use of resources haven't worked for me. Thanks!
First try, worked awesome, image looked great, 2nd image, error: "The size of tensor a (48) must match the size of tensor b (96) at non-singleton dimension 3" Any Ideas?
My new go-to for upscaling. Great output for the speed and good balance between control and simplicity. I am mainly working with graphic novel styled 4-color illustrations instead of photography or cinematic, and was able to get what I needed by including some style guidance in the prompts.
I wasn't able to use fp8_unet as got this: error occurred when executing SUPIR_sample: mat1 and mat2 must have the same dtype, but got Float and Float8_e4m3fn
I've been trying this method and I always get slight multicolor blotches/artifacts over the entire image. Like the upscale is good, but it makes everything look like it has a rainbow sheen. I can't get rid of it no matter what I try.
Hi, Sebastian. Thanks for video! My image comparer got error at the finish "Error occurred when executing Image Comparer rgthree" how can i fix that? Thank you
What are the minimum RAM and VRAM requirements for this to function? I remember when you first mentioned Supir the requirements were really high, but I have a 4070, so I only have 12 gigs of VRAM. I don't care about speed so much, I just don't want to go through the process of figuring out how to set this all up if it's just gonna wind up impossible to even use it.
You can use it but use it tiled for VAE encode and decode + sampling. Make sure the stride is twice as small as the longer end of your tiles resolution
I'm having issues installing "Comfyroll Studio" and "KJNodes for ComfyUI" This doesn't work without them unless I can swap those nodes with something else. They don't look very important. It's mainly the Colormatch node as far as I can tell.
when I try supir.json and click on "Queue Prompt" I get "Error occurred when executing CheckpointLoaderSimple: 'NoneType' object has no attribute 'lower'", do I need to update something?
SUPIR_sample Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype. computer: 128gb ram m2 mac ultra............ didnt get that error on the PC though
I went into the manager to install custom nodes, the one I can't find is "Upscale by X Factor" - Didn't find it with a google search either. Is there a trick to finding that node? Can I switch it for a different upscaler? Thanks!
Is this method better than ultimate sd upscale to upscale photos of human characters with realistic skin? Because I'm using the Supir but I notice that skin blemishes and moles get removed sometimes when I use it. Is there a better setting in Supir I should be using?
Download the work flow but having troubles with it. Error occurred when executing SUPIR_conditioner: 'captions'. I thought the captions input node was optional and I have no idea how to correct that.
The UI would be much neater if there was one node "parameters" where you can connect the desired values from different nodes to one place - and then hide the other nodes. Hope someone implements this feature.
Up to this day I can’t find an upscaler in the level of “Creative Upscaler” used in Night cafe online social AI platform, I wonder is it open source, as many of their models or it’s a dedicated 3rd party upscaling service ?
Aren't you the author? 😅 My understanding is it's similar to an img2img pass or a refiner pass, however people want to call it. I haven't done a side by side comparison, but I would expect mcboaty to have more changes in the final image, making mcboaty (and similar sd ultimate upscale upscalers) to provide beautiful looking results, whereas Supir may be more close to the original input. Do you agree with that assessment?
@@sebastiankamph yes, I am. As other upscalers like SUPIR or USDU, McBoaty is an Upscaler/refiner on a per tile base allowing you to increase the output size with more control over the Denoise. Actually for the Denoise McBoaty v5 will give you that level of control but it is still in heavy dev as we speak
Wasn't able to get this one to work. First it stopped at the conditioner with "'NoneType' object is not callable". So I tried switching models and after that nothing worked. Seems like a lot of people are experiencing this with Supir so it isn't your workflow. Which looks very solid.
@@HistoryIsAbsurd theres a few different spots to change the sizes, even when going really low it doesnt help, is there a forum with more info on this somewhere?
6GB GTX 1060, System I5 32GB RAM. I can get a max of 2X scale and got to use fp16 diffusion_dtype. Q/F checkpoint does not give much difference for my sightseeing. Great vid, thanks.
thanks for the workflow! i got errors in thinkdiffusion, both while installing custom nodes & the SUPIR model. however I dropped it into openart and it worked great
Nice workflow! Is there a way you can upgrade this workflow to make it more of a creative upscale? Like magnificai. Also... batch processing would be great. Many thanks in advance!
Ive tried this and its honestly not great, I've had more success using clarity-upscaler with CCSR. Its also a hell lot slower than other models to boot
Thanks Sebatian very helpful. Still trying to get my head around the nodes you introduced outside of Supir so I can modify my own workflow. Find the right settings as you point out is the trick here. I've found it works really well with very graphic 2D style images. I've yet to find a good mix for painted skin tones, everything is over baked. Has great potential so thanks again. As to the crazy into...well why not! 😊
There was this viral vid that everyone was talking about. AI Warper done joker walking out at a concert. The artist was yachty. Theres a tutorial on it and if u can help with config & overview i'm happy to pay. Thanks. L
@@Omsip123 i actually do. 15 months of it... i think i just been avoiding this, since there is so many websites like krea - but i know i should give it a try
It's extremely amazing how your simple workflow has given me the results that others with hundreds of nodes and models and a huge use of resources haven't worked for me. Thanks!
If you see too many nodes, run!
10/10 Just wanted to say this production quality is setting you apart, quick well lit cuts straight to the point and youre smashing it :D
How very kind of you, thank you so much :) ⭐
dude, the lighting, the editing, you've improved so much, congratulations.
the chair still gives me stephen hawking vibes.
stephen hawking vibes 🤣
Wow, thanks! Really glad you enjoyed it. Exploring in what direction to take the channel and videos
I WAS WAITING FOR AGES HERE FOR THIS ONE, FINALLY
I was a bit sceptical at first, but you earned a new subscriber. No bullshit explanations and just what we need to know. Thank you!
Thank you, very kind :)
I still live in the 1890s and use A1111 Is there a version of Supir for me?
No, you're just normal.
ComfyUI is disgusting to use and look like shit.
@@ADZIOO skill issue
First try, worked awesome, image looked great, 2nd image, error: "The size of tensor a (48) must match the size of tensor b (96) at non-singleton dimension 3" Any Ideas?
I've got the same error
Simply change sampler from TiledRestore to Restore. I personally perfer RestoreDPMPP2MSampler to RestoreEDMSampler.
My new go-to for upscaling. Great output for the speed and good balance between control and simplicity. I am mainly working with graphic novel styled 4-color illustrations instead of photography or cinematic, and was able to get what I needed by including some style guidance in the prompts.
Happy to hear it working well for you when adapting the prompts!
I wasn't able to use fp8_unet as got this: error occurred when executing SUPIR_sample: mat1 and mat2 must have the same dtype, but got Float and Float8_e4m3fn
Same and i cant find that error anywhere else
did you find any solution?
why is it stop loading for me at SUPIR_Conditioner does anyone know? RuntimeError 9
I updated everything
is there a way for this to be available in a1111?
Just 2 questions: 1-can i use an SD15 or must be SDXL? 2-Can it run on a CPU-only ComfyUI?
Can u please share workflow somewhere else instead of Patreon? Patreon is not available in many countries...
Ping me on Discord
I've been trying this method and I always get slight multicolor blotches/artifacts over the entire image. Like the upscale is good, but it makes everything look like it has a rainbow sheen. I can't get rid of it no matter what I try.
Hi, Sebastian. Thanks for video! My image comparer got error at the finish "Error occurred when executing Image Comparer rgthree" how can i fix that? Thank you
Totally loved the opening. :) Thanks so much for this. Been waiting for a SIMPLE workable workflow for this. Glad to be a patreon member!
Awesome, thank you! And glad to hear you liked my little cinematic thing I did ☺️💫
why my pc said reconneting when queue?? help
What are the minimum RAM and VRAM requirements for this to function? I remember when you first mentioned Supir the requirements were really high, but I have a 4070, so I only have 12 gigs of VRAM. I don't care about speed so much, I just don't want to go through the process of figuring out how to set this all up if it's just gonna wind up impossible to even use it.
You can use it but use it tiled for VAE encode and decode + sampling. Make sure the stride is twice as small as the longer end of your tiles resolution
that's really cool!
It would be interesting to see what happens when you add an image to text node and use that as the input description
Is there a way to remove the noise/artifacts? Like on 6x upscale I get squares
I'm having issues installing "Comfyroll Studio" and "KJNodes for ComfyUI"
This doesn't work without them unless I can swap those nodes with something else. They don't look very important. It's mainly the Colormatch node as far as I can tell.
when I try supir.json and click on "Queue Prompt" I get "Error occurred when executing CheckpointLoaderSimple:
'NoneType' object has no attribute 'lower'", do I need to update something?
SUPIR_sample Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.
computer: 128gb ram m2 mac ultra............ didnt get that error on the PC though
hey sebastian, what do you recommend now? looking for a simple workflow that is similar to krea upscale and enhance.
help :c in the pink box I get a total black image, what can I do :C
I went into the manager to install custom nodes, the one I can't find is "Upscale by X Factor" - Didn't find it with a google search either. Is there a trick to finding that node? Can I switch it for a different upscaler? Thanks!
You can upscale with any upscaling node. On mobile atm so can't check the node name I used.
Is this method better than ultimate sd upscale to upscale photos of human characters with realistic skin? Because I'm using the Supir but I notice that skin blemishes and moles get removed sometimes when I use it. Is there a better setting in Supir I should be using?
Download the work flow but having troubles with it. Error occurred when executing SUPIR_conditioner: 'captions'. I thought the captions input node was optional and I have no idea how to correct that.
The UI would be much neater if there was one node "parameters" where you can connect the desired values from different nodes to one place - and then hide the other nodes. Hope someone implements this feature.
Interesting video very good explaining, not boring at all. Continue man
Thank you, very kind :)
wow creative talent in the edit here SIr! ANd of course good knowledge as ever. Thank you.
So nice of you, thank you ☺️💫
Where does the upscale part happen?
Or is it just meant to upscale it before it run through SUPIR Sampler, is that correct?
I have an error. mixed dtype: expect parameter to have scalar type of float
is there anyway to get this effect but in a series of image frames to make a video?
Error occurred when executing SUPIR_sample:
The size of tensor a (32) must match the size of tensor b (96) at non-singleton dimension 3
What is this! That intro is AMAZING!
Thank you buddy! I tried something new, glad you liked it 😁 Saw you got a new job, congratulations! All good?
@@sebastiankamph it's genuinely good! And thanks, those are just part time roles but they're definite first steps.
Error occurred when executing SUPIR_sample:
mat1 and mat2 must have the same dtype, but got Float and Float8_e4m3fn .
How to rectify this error?
did you resolve it?
I cannot get it to go beyond the SUPIR Conditioner. I get error " Error occurred when executing SUPIR_conditioner:
'NoneType' object is not callable"
This was a really good help for me, thank you very much!
Thank u for this amazing tutorial, i love how the editing qualities of your videos is upscaled too.😍😍👍👍
Happy to hear it! 😁💫
Up to this day I can’t find an upscaler in the level of “Creative Upscaler” used in Night cafe online social AI platform, I wonder is it open source, as many of their models or it’s a dedicated 3rd party upscaling service ?
wow ! how to link checkpoint files from comfyui, a1111, fooocus, etc ? i keep downloading multiple , rip my storage
Hi, have you tried McBoaty V3 to compare against SUPIR, I would love to have your insight on it
Aren't you the author? 😅 My understanding is it's similar to an img2img pass or a refiner pass, however people want to call it. I haven't done a side by side comparison, but I would expect mcboaty to have more changes in the final image, making mcboaty (and similar sd ultimate upscale upscalers) to provide beautiful looking results, whereas Supir may be more close to the original input. Do you agree with that assessment?
@@sebastiankamph yes, I am. As other upscalers like SUPIR or USDU, McBoaty is an Upscaler/refiner on a per tile base allowing you to increase the output size with more control over the Denoise. Actually for the Denoise McBoaty v5 will give you that level of control but it is still in heavy dev as we speak
Thanks Seb! I'm really leaning into learning how to use Comfy UI now gaaaaaaaaaaaaaah
Where can I buy your pc or is it custom built?
the workflow seems to be broken on updated comfy
Wasn't able to get this one to work. First it stopped at the conditioner with "'NoneType' object is not callable". So I tried switching models and after that nothing worked. Seems like a lot of people are experiencing this with Supir so it isn't your workflow. Which looks very solid.
That error means you're trying to call and object outside of the size of the current tuple. Try changing the tiles size of your sampling.
@@HistoryIsAbsurd theres a few different spots to change the sizes, even when going really low it doesnt help, is there a forum with more info on this somewhere?
6GB GTX 1060, System I5 32GB RAM.
I can get a max of 2X scale and got to use fp16 diffusion_dtype. Q/F checkpoint does not give much difference for my sightseeing.
Great vid, thanks.
Is it somehow possible to useultraSharp resampler instead of Lanczos?
Yes! Connect a node loading that going into the upscaler. Will smooth out the results a little
thanks for the workflow! i got errors in thinkdiffusion, both while installing custom nodes & the SUPIR model. however I dropped it into openart and it worked great
what are your PC specs: RAM + VRAM???
64gb ram, rtx4090 with 24gb vram
How did you get the text in the image compare? The text before and after over each of the images?
I think it's called cr watermark. Check the nodes behind that box.
@@sebastiankamph I found another compare node that writes out a text but shows the images side by side. Rather than one over the other
Whats the extension for the usage stats?
Crystools. I made a video on it
@@sebastiankamph thank you. will take a look.
Nice workflow! Is there a way you can upgrade this workflow to make it more of a creative upscale? Like magnificai. Also... batch processing would be great. Many thanks in advance!
More s noise and ddm eta less controlnet
Thank you, great tutorial and really pretty workflow :)
You're very welcome! And thank you 😊🌟
I tried to use it 1:1 so just as a denoiser and it had a singleton error. :/
Thank you so much dear sir! Great Workflow, nicely explained
So nice of you! Most welcome.
Thanks, works great on my RTX 4060TI 16gb.
03:36 "i will link this one in the comments bellow" (doens't link anywhere)
Do I can use it like an API?
Yes, you download the JSON in the API format and can run it
how to connect Supir with controlnet or ipadapter as a workflow in comfyui ?
You dont have to SUPIR comes with controlnet inside of it just set it to 1.00 on both start and end but also dont have alot of noise and dd eta
Hey thank you for the video. I think you forgot to comment the link to lightning model. ( 03:35)
Ive tried this and its honestly not great, I've had more success using clarity-upscaler with CCSR.
Its also a hell lot slower than other models to boot
great workflow thanks,
Happy to help!
0:18 what do get if you set a puppy on fire?
A lifetime ban from entering the pet store
Thanks very much. Very cool!
14:50 "it's a small image... 1024x1024"
Windows XP has left the chat 😂
Thank you very much for the wonderful diagram, it works great.
You are welcome!
Is this better than magnific?
When used correctly and tied with more advanced settings, yes. Magnific runs multidiffusion tiles as far as I know
Thanks Sebatian very helpful. Still trying to get my head around the nodes you introduced outside of Supir so I can modify my own workflow. Find the right settings as you point out is the trick here. I've found it works really well with very graphic 2D style images. I've yet to find a good mix for painted skin tones, everything is over baked. Has great potential so thanks again.
As to the crazy into...well why not! 😊
Happy to help! Hope it'll get you some great images. Thank you for the comment 😊🌟
Simple wow bro!
✨👌😎😯😎👍✨
Thanks ✌️
I try to use this, and my PC immediately crashes.
Dad jokes equal insta-likes from me XD
Also, nice video, as always! :D
Glad you like them! And thanks for the likes and comments always 😊💫
this is excellent
Thanks a lot! Greetings from Uppsala!!!
What!? Are you travelling?
Thanks
Thank you
☺️💫
it made my image worse?
There was this viral vid that everyone was talking about.
AI Warper done joker walking out at a concert.
The artist was yachty.
Theres a tutorial on it and if u can help with config & overview i'm happy to pay.
Thanks.
L
Everything to say he has a 4090...
👋
Oh why hello, good to see you again!
@@sebastiankamph
🫂
Yikes, :37 secs I was out! Ha toooo confusing for me bro
It's setup so it's almost oneclick. You don't need to learn the settings but I go over them :)
Greetings from Russia! Thank you for the tutorial!
You are welcome!
mumbling voice with loud bing bong bing music 👎👎
ai cant teach me how to set this up i bet
It can
You don’t know AI then….
@@Omsip123 i actually do. 15 months of it... i think i just been avoiding this, since there is so many websites like krea - but i know i should give it a try
Not really
Says the AI consultant 😂😂
Thank you