i have 32gb and with rtx 3090 getting this error with xl and for real both: SUPIR_sample The size of tensor a (36) must match the size of tensor b (96) at non-singleton dimension 3
Thank you! I downloaded and tried many Supir workflows, and this is the only one that works great with amazing quality. This workflow is very easy to understand and well-organized. I used the RealVisXL V50 SDXL model on my 3060 with 12GB VRAM, and it took 4 minutes. The quality is definitely worth the wait.
@@Jockerai In his previous video. (He speaks about it in the beginning.) TBH this wf is nice but there are better ones. (of course there is always better). So I am not putting this one down just when I tested it stuff like people's faces it will change too much.
Interesting. Isn't this the old Supir model that we've had for more than a year now though? I already have the SUPIR--v0F and SUPIR-v0Q models installed. Are these some new higher quality models? Judging by their names, they are just pruned versions (reduced file size) of the original SUPIR models, or am I wrong? Can you do a comparison between these 2 models and the old ones from a year ago to see the difference in quality?
Can you use a Lora with this Upscaler? Or is that just not possible? I like XLabs Flux Realism LoRA, and this upscaler changes the vibe of those images dramatically, I'd like for it to keep the essence intact
I didnt test any lora for this workflow but I think it's better to use img2img for that. just befor upscale make the image the way you like then upscale it.
@@Jockerai Although it messes up my female model faces. Specially when it's a long shot instead of a close-up. How could I fix that? Other than that it's amazing
@@Jockerai Can I connect a face detailer to it at the end? I can't seem to be able to find a way to connect the positive and negative from the FaceDetailer node because that sampler doesn't have those outputs
@@samu7015 consistent character faces is common issue for everyone specially when face is isn't close enough. The best way to make faces unchanged is using IP Adapter which I will make a video for that soon
@@Jockerai Intel(R) Core(TM) i9-14900K 3.20 GHz, RAM 64,0 Go, Windows 11 Professionnal, Nvidia Geforce RTX 4090. Let me know if you need something else, many thanks
I'm going to upload a complementary video for this and another tips for this which I forgot mention in this video😎. But for now just change the "sampler" in "SUPIR sampler" to "RestorEDMsampler" it will solve your error.
@@AInfectados 2 workflows will work for you. It is not related to GPU, the difference between workflows is about SDXL models. If you use a SDXL model which needs VAE you have to use second workflow. But there are some advises for good GPUs and weak GPUs that I forgot to add in my video. I'm going to upload a complementary video for this.
just go to manager and click "Install Missing custom nodes" and then install all of nodes that are found then restart ComfyUi and the errors will be gone
@@Jockerai I'm going to start from scratch and follow your comfy install. I have python installed in general on the system do you think it's best just to have portable standalone build and not have python installed for a noob to avoid conflicts?
@@Jockerai oh okay, because in the video you said version 3, and I thought something's wrong lol, anyways, thanks for the tutorial! great content, and this resizer is fire, bye bye krea :D
@@Jockerai by the way, the workflow works like a charm, however on certain sized images, it won't do, and I can't tell what sizes it accepts, any idea about those limitations?
It works on Mac M1 and M3. I'm using it for many months now. Note: everyone knows NEVER to by Apple even-numbered gear (2,4,6,8,etc,); always odd-numbered.
No good video to be honest. Supir for ComfyUi is non commercial only so not very useful for most use cases. It is only for research only. You mentioned commercial sites and doing wrong claims ("completely free,...") which is not correct.
click the link below to stay update with latest tutorial about Ai 👇🏻 :
www.youtube.com/@Jockerai?sub_confirmation=1
i have 32gb and with rtx 3090 getting this error with xl and for real both: SUPIR_sample
The size of tensor a (36) must match the size of tensor b (96) at non-singleton dimension 3
Thank you! I downloaded and tried many Supir workflows, and this is the only one that works great with amazing quality. This workflow is very easy to understand and well-organized. I used the RealVisXL V50 SDXL model on my 3060 with 12GB VRAM, and it took 4 minutes. The quality is definitely worth the wait.
You're welcome bro✨😉
that fulx did not use controlnet tile.
@@zGenMedia which Flux?
@@Jockerai In his previous video. (He speaks about it in the beginning.) TBH this wf is nice but there are better ones. (of course there is always better). So I am not putting this one down just when I tested it stuff like people's faces it will change too much.
Just amazing, the best upscaler I tryed, thanks a lot for sharing all this 👍
Adding *Florence* node to get the description of the image automatically would be great.
That what I was thinking, just change the Supir text to an input and attach Florence node with a load image! Awesome idea.
@@A_Little_O_Everything Can you share the workflow with that modification?
@@AInfectadosthat explanation is really straight forward
Thank you for sharing. Will it work with 3D or Painting style too?
you're welcome. Yes sure
Interesting. Isn't this the old Supir model that we've had for more than a year now though? I already have the SUPIR--v0F and SUPIR-v0Q models installed. Are these some new higher quality models? Judging by their names, they are just pruned versions (reduced file size) of the original SUPIR models, or am I wrong? Can you do a comparison between these 2 models and the old ones from a year ago to see the difference in quality?
@@bgtubber I'm not sure but I think it is different. I will search for it to make sure
Благодарю за вашу работу, все отлично.
English please
hello... i tried this tutorial but comfyui gives this error: Cannot execute because a node is missing the class_type property.: Node ID '#15'
@@KNO1973 watch the next video. Link in the description
Can you use a Lora with this Upscaler? Or is that just not possible? I like XLabs Flux Realism LoRA, and this upscaler changes the vibe of those images dramatically, I'd like for it to keep the essence intact
I didnt test any lora for this workflow but I think it's better to use img2img for that. just befor upscale make the image the way you like then upscale it.
@@Jockerai Although it messes up my female model faces. Specially when it's a long shot instead of a close-up. How could I fix that? Other than that it's amazing
@@Jockerai Can I connect a face detailer to it at the end? I can't seem to be able to find a way to connect the positive and negative from the FaceDetailer node because that sampler doesn't have those outputs
@@samu7015 consistent character faces is common issue for everyone specially when face is isn't close enough. The best way to make faces unchanged is using IP Adapter which I will make a video for that soon
*What is the difference between forrealxlV10_v10 and forrealxlV10_v03?*
simply V1.0 is an update from V0.3. V1.0 doesn't need VAE and it is more optimized and has better train
how can I miss a lot node after I loading your workflow
@@neoviva please watch next video which it's link is in description
Cannot execute because a node is missing the class_type property.: Node ID '#15'
watch the next video . link in description
same here, could you solve it?
I'm getting this error: SUPIR_sample
Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.
I upload a new video to solve issues in Supir. please check it and do what I said. It will probably solve your issues
@@Jockerai I'll see it. Thank you!!
@@RubenAlejandroTrejoSegura you're welcome
I solved it by turning off the option "fp8_unet" setting this to "false".
@@RubenAlejandroTrejoSegura thanks for sharing this
Thanks for the video. I get this error, can you please help?
"No operator found for `memory_efficient_attention_forward` with inputs: query..."
please check the new video that I have just uploaded. Do whatever I recommend in that video.
@@Jockerai Yes, I watched the video, but my issue doesn't seem to be addressed. Can you help me?
@@graphilia7 what is your system config?
@@Jockerai Intel(R) Core(TM) i9-14900K 3.20 GHz, RAM 64,0 Go, Windows 11 Professionnal, Nvidia Geforce RTX 4090.
Let me know if you need something else, many thanks
@@graphilia7 github.com/AUTOMATIC1111/stable-diffusion-webui/issues/7973 maybe this can help
RuntimeError: The size of tensor a (32) must match the size of tensor b (96) at non-singleton dimension 3
I'm going to upload a complementary video for this and another tips for this which I forgot mention in this video😎. But for now just change the "sampler" in "SUPIR sampler" to "RestorEDMsampler" it will solve your error.
I've got the same problem. The error was the side lengths of the image must be a must be a multiple of 64. Worked for me.
@@fittersitter So what number did you use. Sigh...
@@zGenMedia I change sampler_tile_size from 768 to 512 and it got rid of this error. hope this helps
@@CMPCr8vDgs it work yes but the tiles are visible now in my result image, the fix for me was to use RestorEDMsampler instead of the default one
But where is the image with the workflow, i don't see any link....
Check the description.
- Download Supir workflow
@@Jockerai Thx!
@@Jockerai Which one of the 2 workflows you think will be better for me?
I have a 4070 TI Super 16gb.
@@AInfectados 2 workflows will work for you. It is not related to GPU, the difference between workflows is about SDXL models. If you use a SDXL model which needs VAE you have to use second workflow. But there are some advises for good GPUs and weak GPUs that I forgot to add in my video. I'm going to upload a complementary video for this.
I am getting error, plz help :- Cannot execute because a node is missing the class_type property.: Node ID '#15'
just go to manager and click "Install Missing custom nodes" and then install all of nodes that are found then restart ComfyUi and the errors will be gone
@@Jockerai I installed all custom nodes, NO NODE is RED, but after queue prompt, it gives this error,plz help
@@ParthKakarwar-b7j what error you get?
@@Jockerai Actually,my comfy was not updated,it solved the error after that,Thank u,😇
I was really Hoping it was going to work , the cmd would just crash out on both workflows When it was doing the tiling. 8vram Anybody got any ideas
plz tell when you exactly get the crash?
@@Jockerai I'm going to start from scratch and follow your comfy install. I have python installed in general on the system do you think it's best just to have portable standalone build and not have python installed for a noob to avoid conflicts?
@@WallyMahar yes try this and see what happens
I'm not sure if you noticed, but version 1.0 is newer than version 0.3 ...
Yes I knew that, because I had downloaded V0.3, I didn't want download 13G again and did my work with older version. but you can use v1.0 instead
@@Jockerai oh okay, because in the video you said version 3, and I thought something's wrong lol, anyways, thanks for the tutorial! great content, and this resizer is fire, bye bye krea :D
@@Jockerai by the way, the workflow works like a charm, however on certain sized images, it won't do, and I can't tell what sizes it accepts, any idea about those limitations?
@@nickmack7282 Yes, that mistake wasn’t on purpose🤩
@@nickmack7282 I've explained everything in the next video, make sure to check it out.
Does it work with Flux?
No it doesn't
we need it for mac m2 users
I didnt find a free usable way for mac, I will try to
It works on Mac M1 and M3. I'm using it for many months now.
Note: everyone knows NEVER to by Apple even-numbered gear (2,4,6,8,etc,); always odd-numbered.
No good video to be honest. Supir for ComfyUi is non commercial only so not very useful for most use cases. It is only for research only. You mentioned commercial sites and doing wrong claims ("completely free,...") which is not correct.
This still is not it ngl.... I really hope to see an actual ai that can restore quality not reimagine quality..
Watch my first video in my channel