I also have a 3060 with 12gb Vram. You can run flux dev fp8 22gb variant. But it will take time and will slow down your pc movements too. Don’t even think about upscaling with FP8. I will recommend stick with the turbo lora’s or GGUF models. Or generate low res image with full fp8 and upscale with turbo with 8 step or GGUF. and thanks for the compliment 😬
@@aijigyasa its failing to build the insightface wheel no matter what I do. was reading somewhere the non-portable version does it without probs (maybe I have a path problem) will try tomm
@@MrDebranjandutta try installing insightface with your python version. Steps to Follow: Check the Python version in ComfyUI: 1. Go to the ~/comfyui/python_embedded folder and double-click on python.exe. Note the version of Python (e.g., 3.09, 3.10, 3.11, or 3.12). 2. Now open this link(in description) - github.com/Gourieff/Assets/tree/main/Insightface and download the same version as python. and move that file to comfyui folder where run_nvidia_gpu batch file exist. 3. now right click on it and copy as path. 4. right click anywhere in the folder and open in terminal and paste this command - .\python_embeded\python.exe -m pip install [paste path here] onnxruntime and hit enter. that's it. let me know if it fixes your problem.
@@aijigyasa Yes, i have watched your video till the end. If you want and have the time can you make a flux workflow to create ultra realistic professional photos (not the plastic skin look with bad stare) based on uploaded photo with positive and negative prompts: 1. Consistency in character (currently i am working with fooocus and i have a female model i use in front of a black backdrop and when i want to use her i simply drop her photo and generate photos) 2. that fix the face that match the uploaded photo 3. next to normal generated photo an additional upscaled photo using the ComfyUI-Flowty-LDSR (currently one of the best for upscaling)
i coulld not get it working on windows, but worked like a charm in ubuntu. I wonder if the upscaling nodes are necessary, considering default upscale value is 1.20. As it is it's quite slow, and upscaling just adds to the time
I think the problem is with Ultimate SD upscale. comfyui started rolling out vae decode tile with an overlap option. Wait for them to fix it. Meanwhile, you can turn off the Vae decode option at the bottom of Ultimate upscaler.. That should fix the problem.
it'd be funny 2-3 years from now yall get sued for using celebrities and that was part of the ai scam hahahahahahhahaha Oh man it'd be a laugh to watch all the people cry and act like they didnt know.... Pretty cool times we live in
@@Romeo615Videos this video is for educational purposes only. I used a celebrity face bcuz it’s easy to find high quality image. And it’s public. These images were used for all sorts of work memes or small ads. So i think we will be fine. But thanks for showing concern.
Good stuff bro, subsd. Which gpu are you using btw? I have a poor mans 3060 can I use the full dev-fp8 model?. Like your goofball style. Keep at it
I also have a 3060 with 12gb Vram. You can run flux dev fp8 22gb variant. But it will take time and will slow down your pc movements too. Don’t even think about upscaling with FP8. I will recommend stick with the turbo lora’s or GGUF models. Or generate low res image with full fp8 and upscale with turbo with 8 step or GGUF. and thanks for the compliment 😬
oh my god.... thx as always.
No worries. Keep Supporting.
@@aijigyasa its failing to build the insightface wheel no matter what I do. was reading somewhere the non-portable version does it without probs (maybe I have a path problem) will try tomm
@@MrDebranjandutta try installing insightface with your python version.
Steps to Follow:
Check the Python version in ComfyUI:
1. Go to the ~/comfyui/python_embedded folder and double-click on python.exe.
Note the version of Python (e.g., 3.09, 3.10, 3.11, or 3.12).
2. Now open this link(in description) - github.com/Gourieff/Assets/tree/main/Insightface
and download the same version as python. and move that file to comfyui folder where run_nvidia_gpu batch file exist.
3. now right click on it and copy as path.
4. right click anywhere in the folder and open in terminal and paste this command - .\python_embeded\python.exe -m pip install [paste path here] onnxruntime
and hit enter. that's it. let me know if it fixes your problem.
"She looks exactly like her ... HMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMMM" 😭😭😭😭😭😭🤣🤣🤣🤣🤣🤣
You loved the bit. I can tell by looking at your face. 🤣🤣🤣🤣🤣🤣
@@aijigyasa ah man you made me laugh
@ I am glad to know.
@@aijigyasa Yes, i have watched your video till the end. If you want and have the time can you make a flux workflow to create ultra realistic professional photos (not the plastic skin look with bad stare) based on uploaded photo with positive and negative prompts:
1. Consistency in character (currently i am working with fooocus and i have a female model i use in front of a black backdrop and when i want to use her i simply drop her photo and generate photos)
2. that fix the face that match the uploaded photo
3. next to normal generated photo an additional upscaled photo using the ComfyUI-Flowty-LDSR (currently one of the best for upscaling)
@ sure will look into it.
i coulld not get it working on windows, but worked like a charm in ubuntu. I wonder if the upscaling nodes are necessary, considering default upscale value is 1.20. As it is it's quite slow, and upscaling just adds to the time
great. Upscaling is optional. I used it to refine some minor details; that's it.
Using your workflow and exact models/loras, I got this error:
VAEDecodeTiled.decode() missing 1 required positional argument: 'overlap'
I think the problem is with Ultimate SD upscale. comfyui started rolling out vae decode tile with an overlap option. Wait for them to fix it. Meanwhile, you can turn off the Vae decode option at the bottom of Ultimate upscaler.. That should fix the problem.
the final image its so blury
Increase the steps. Use Lora detailers; put blurry in negative when upscaling. that should fix it
it'd be funny 2-3 years from now yall get sued for using celebrities and that was part of the ai scam hahahahahahhahaha Oh man it'd be a laugh to watch all the people cry and act like they didnt know.... Pretty cool times we live in
@@Romeo615Videos this video is for educational purposes only. I used a celebrity face bcuz it’s easy to find high quality image.
And it’s public. These images were used for all sorts of work memes or small ads. So i think we will be fine. But thanks for showing concern.