Notice: If you encounter an error for the SAM model; download the SAM model I have provided and place it in the SAMS folder under the Model Directory in your comfy install.
LowVRAM Compatible Workflow (Remastered)
This is a workflow intended for LowVRAM Users like myself, but can be used with Higher VRAM Cards as well.
SPECIAL THANKS TO : @Jadewind for helping me identify issues with the last version that I missed! Killer job friend!
Tested on a 3050Ti 4GB Mobile GPU
Tested on all SDXL Resolutions
Pre-Configured
Download links included for Detailer and Upscaler Models
Built in OoM (Out of Memory) "Buffers" for convenient restart points.
Full ADetailer type Pipeline
Dual Upscalers (CR and Latent)
Includes "Anything Everywhere" Nodes for a clean UI experience (No Spaghetti)
Efficiency Nodes for reduced node clutter
Group Notes for explanation and configuration purposes.
Description
This version fixes a number of issues associated with the workflow that I overlooked before release.
CR Upscale Buffer was not connected to the Latent Upscale Nodes and was passing straight to Save Node. (Fixed)
Incorrect Link for one or more Detailer Models (Fixed)
Included directions for placing Detailer models (added)
Double checked workflow for additional issues that may have been missed (Verified)
If you encounter any issues with the updated workflow please contact me directly via Chat or leave a comment on the page.
Thank you and enjoy!
FAQ
Comments (14)
The workflow is quite interesting.
The Path AddDetailer Yolo´s:
Data\Packages\ComfyUI\models\ultralytics
One thing I don't understand.
After the wathermark it saves the image where it should be in a resolution of 1248x1824 which is ok.
But then it does an upscale which takes about 6 minutes:
IterativeLatentUpscale[Final]: 2496.0x3648.0 (scale:2.00) but I can't find this file anywhere?
Hey there. So, I guess I somehow over looked this before I released it. You'll need to connect the images output on Buffer-4 to the images input on the Lucy Sharpen Node. Sorry about that. and thanks for bringing it to my attention.
@DogFace That was exactly the mistake, and now Buffer 4 is also filling up. The quality is fantastic.
This is one of the cleanest and best workflows I have ever seen. I would write in the instructions where people should copy the yolos. I only figured it out because I already had the hand Yolo and could search for it. I would never have found the memory path :)
One more question, since I have an NVIDIA with 16GB, could I adapt the workflow to it? Or would you build another version of the workflow with High VRam?
And the last question, the watermark with the text is great. However, I have a signature that a company designed for me. Is there a way to use a real watermark instead of text, e.g. in a fixed size at the bottom right? Sorry for the many questions... but now there is a Buzz donation :)
@Jadewind Thanks for the kind words and the donation! I'm happy that you're enjoying it.
To your questions:
1. The workflow, in it's default form, is built to work for 4-8GB Workflows. However, yes it can be adapted to work with higher VRAM Cards such as your 16GB Card. The one area that you can play around with is the Latent Upscale. Specifically the number of passes it performs. 4 Passes was the most that I could squeeze out of my 4GB card. With 2 Passes offering the most consistency. But with 16GB VRAM I'd gather you could probably get 6-8 passes if not more.
I'd love to make a workflow that better caters to cards of your capability, but alas, I'm working with a 4GB VRAM card in my Laptop (3050Ti). So it's not possible for me to create such a workflow and test it properly for release. I'm saving up to build a rig but that's slow going as I'm also saving up for my wedding as well.
2. As far as the watermark is concerned; I do know that there are some nodes out there that can do what you're wanting to do. But because of the VRAM constraints I listed above, I'm not able to test them and give my opinion on them one way or the other. In my case I'd have to rely on a program like Krita, Gimp, or Photoshop to accomplish that. I'd imagine that since you had it made by a pro that it's a transparent/vector image. So doing it that way would definitely work out for you.
3. Instructions: You're correct here. Since you've brought the Node issue with the Latent upscale to my attention; I'll be rereleasing the workflow here in a day or so. I'll include some directions on where the files should be placed. I was hesitant on doing this previously simply because some users point Comfy to their A1111 and Forge install folders for models and such. But I'll just include instructions for both.
Thanks again for the kind words and donation. It is much appreciated and motivates me to keep working on new workflows for release!
@DogFace I have to thank you my friend. I will google through the Watermark Nodes :)
Thank you for the very detailed answer.
Thanks for the Watermark tip, I have now built a complete workflow into yours and I am fascinated.
@Jadewind Very Nice! I really need to get my signature game up! Looks great!
There is one problem that I still haven't got to grips with. With photos that are very close to the face, we call them close-ups, the workflow does very strange things. Even if I deactivate almost all of the detailers with bypass, the result is not good. The initial image is almost perfect and it only gets worse as the workflow continues. Once you get a certain distance from the face, the problem disappears immediately.
Hmmm... Can you chat me some examples of the issue? If so: Send me the initial Gen, the CR Upscale Gen, and the Latent Gen. Since your bypassing the Detailer Pipeline I'm gathering that you haven't changed any settings there.
If you have changed settings for the CR Upscale and the Latent Upscale please send me pics of those settings as well.
I have personally Gen'ed a lot of headshots and Close ups with this workflow and haven't had an issue like that.
Also, what model are you using so I can do some independent testing.
Getting this error:
Error occurred when executing SAMLoader: 404 Client Error: Not Found for url: https://github.com/ultralytics/assets/releases/download/v8.2.0/sam_vit_b_01ec64.pth File "/home/studio-lab-user/ComfyUI/execution.py", line 151, in recursive_execute output_data, output_ui = get_output_data(obj, input_data_all) File "/home/studio-lab-user/ComfyUI/execution.py", line 81, in get_output_data return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True) File "/home/studio-lab-user/ComfyUI/execution.py", line 74, in map_node_over_list results.append(getattr(obj, func)(**slice_dict(input_data_all, i))) File "/home/studio-lab-user/ComfyUI/custom_nodes/ComfyUI-YOLO/nodes.py", line 296, in load_model response.raise_for_status() # Raise an exception if the download fails File "/tmp/venv/lib/python3.10/site-packages/requests/models.py", line 1021, in raise_for_status raise HTTPError(http_error_msg, response=self)
interesting... It looks like the creator has done some restructuring on their Github page 17 hours ago. I'm not sure what this means for the models as I can't seem to locate the model in the file structure at the moment. I'll give it a day or so and check back in on the page to see if there is any change. For the time being it seems that the node is broken for users that are downloading the SAM models for the first time.
I'll upload the SAM Model here so you and others can download it.
I have added a download for the SAM model in the versions ribbon. Download that and place it in the SAMS folder under the Models Directory in your Comfy Install. Thanks for bringing this so my attention.