This workflow is intended to be used with the Z-Image Turbo model to generate an image and then refine the image using the GonzaLomo V6.0 Photo XL model. V6.0 will provide the image with much better detailed anatomical features.
Note that it's not required to use the GonzaLomo SDXL model as any photorealistic SDXL model should do a very good job at refining the Z-Image image.
The different detailers require separate models that should go in your ComfyUI /models/ultralytics/bbox folder. I've uploaded them to this model page here but you can also find these files here and here.
Description
Initial version with multiple detailers.
FAQ
Comments (41)
The refinement will be applied only to the face or all the image?
The upscaler will provide a small amount of refinement to the whole image. That can be increased or decreased by changing the denoise. Then there are multiple detailers that provide more detail to individual areas of the body in the image.
Extremely good! The file names for the detailers are a bit different than the links, but after some straightforward deduction and renaming everything works great. Thanks so much for providing this great workflow!! The easy to understand notes are extremely helpful.
So good... so very, very good... thanks and enjoy the buzz...
Cool idea! I was getting hung up with a "Clip Loader" error... But I needed a ComfyUI update... better now!
this worked for me!
Lol z-image already has detailed anatomical features
Not sure what you mean. Nipples and genitals are definitely missing detail because Z-Image was undertrained in those areas.
@GBRX I personally think that after the base model is realeased we can train loras for nipples and genitals. But you made a great temporary solution for now.
@coolstrad Yeah, this workflow is only meant to be a stopgap until the better trained checkpoints start appearing. Refining should really just add detail, not fix major detail flaws.
I don't understand the thinking behind this, it should be the opposite. SDXL is killing Z-image details!
This is based on my refiner for Flux. I agree that you don't want to affect the z-image details as much as possible so I keep the refiner denoise pretty low. It could even go a bit lower with lower steps. It all depends on the image, seed, etc. There's some trial and error involved.
The main point of the workflow is to provide the nipple/genital detail that z-image currently lacks. Hopefully we get better trained checkpoints soon and that would make this workflow more obsolete.
Any unit can be independently disabled, including the refiner. At the bottom of the diagram, the author has provided such a switch, which is very convenient.
It's better to just decode the z-image output and run some 1-2 detailers on the output for anatomy at low denoise. Complete denoise pass removes the subtle styling of the model.
That's a good idea for another version. Maybe I'll do that this weekend.
I was asked to come up with a workflow on Thanksgiving so I only had time to leverage one of my existing workflows. I appreciate the suggestion though.
I struggled to find a good reliable yolo model for the nipples but this one at a 512 guide and 0.3 threshold and low denoise works well. It's a bit different than the 'nipples' one which likes to make two different nipples. https://huggingface.co/hiramaru999/randomstuff/blob/main/nipple.pt
@tenstrip Thanks for that link! I tried it, and it also seems to do a better job of detecting too!
Can we use loras from other models or do we need to wait for the community to generate them?
I'm not 100% sure but I tried loras from other model architectures and they didn't seem to have any effect.
@GBRX I tried some too - same.
@dcham2310 @GBRX Same for my personal CivitAI created Loras and others... however, in a twist of fate I tried a few celebrity faces (Meatloaf:1.2)... BINGO!! :-P
@GenoMachino Can I ask what you mean by trying faces? Do you mean loras or image2image? If the latter, I have yet to see an image2image workflow and am not sure how to go about making one.
@dcham2310 In the past using a public person's name in the prompt would lean the image generation toward an image that resembles that person. I am in no way supporting or condoning identity stealing/borrowing, etc... just noticed that while generic LORA's don't seem to work with the ZImage checkpoints, public figure names might influence the images.
@GenoMachino OK - thanks, that makes sense. My hopes were in another direction as I'm not interested in celebs but using myself and friends in comical or fantastical scenes.
Thank you for your excellent work. This is the only working workflow that immediately started working on my equipment. In my opinion, everything is thought out and optimally configured in it. The only thing that raises doubts is the relevance of the refiner as implemented in the workflow. It would be great if you could find the time for a similar workflow for I2I.
Thanks so much. I'm actually in the final stages of a V2.0 that I think will be much better. I just quickly threw V1.0 together on Thanksgiving and it didn't get the full amount of testing that I would normally give a release. I totally agree that the full SDXL refiner is unnecessary and can hurt the final image.
Hi friend, I have a question. Why isn't my LoRa working in the workflow? It's not being recognized. Could you help me with instructions on how to get it working with LoRa? Thank you very much.
I haven't had any issues with loras working. I've been creating ZIT character loras and using them in this workflow. All you should have to do is add the lora to the Power Lora Loader and enable it. Is this a publicly available lora I could try?
Closed as not planned
Cannot execute because a node is missing the class_type property.: Node ID '#18'
I'm sorry but that is not enough information for me to know what the problem could be for you. Which version of the workflow are you using- 1.1 or 1.0? At what point in the workflow did you get that error? ie which was node was being executed?
Hey GBRX thank you so much for the workflow and your other work! So helpful. I was curious, do you happen to have a guide or could point me to one in regards to using the FaceDetailer? I'm having some trouble with the nipple detail and would love to know what parameters to adjust to make it work better and whether you could do things like make the areolas larger or smaller. Nipples more erect etc.
You could try adding a lora to the SDXL Detailer Model Setup that would change the nipple style. I'm sure there's no shortage of SDXL nipple loras that you can try but I don't have any specific ones to recommend. The only parameter in the nipple detailer you should really change is the denoise. Making the denoise higher will allow the detailer to change the nipples more.
@GBRX Got it thanks again!
Everytime I run this workflow, it reverts all positive prompts back to your arab model... why?
The workflow file itself contains that prompt. You can change the prompt to whatever you like or leave it empty and then save it locally as a workflow.
@GBRX Thats what the issue is. I've swapped out all those and as soon as i hit run, it reverts back to your prompt. I've swapped it in the clip prompt and the normal prompt.
@ShrubberNi The prompt changes after you hit run?? I've never seen that behavior before and have no idea what that issue could be.
@GBRX I just tested it and if i delete the prompt, as soon as i hit space bar, it reverts back to your prompt. it's so weird
Ah, it doesn't like node 2.0. Works if i revert back
@ShrubberNi Oh ok, I haven't switched over to Nodes 2.0 yet. I tried it once and it did some weird stuff so I'm not going to use it until the issues are worked out.
me too