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
CZXL = Chroma->ZImageTurbo->SDXL
This is a very experimental (and slow) workflow that uses Chroma to create the initial image because I think it has superior composition and color. Also, Chroma provides a wider diversity of faces and body types than ZIT, in addition to much better NSFW options.
The entire image is then passed to ZImageTurbo to be refined on a high denoise, ~0.60, because ZIT provides superior realistic textures.
Finally, there are some optional SDXL refiners to fix the naughty bits that ZIT tends to mess up.
I find that this workflow provides the best of all worlds for me but I wish it wasn't necessary. Hopefully we get much better trained ZIT models and that can take over completely. For now, this is the workflow I'm using on a regular basis.
Please provide me some feedback! I'd like to make this better and/or fix any problems that anyone discovers.
P.S. The quality of the resulting image is going to be totally defined by the different denoise settings. You can play around with the ZIT denoise a lot and go anywhere from 0.15 to 0.75. Also, the SDXL detailers have their own denoise settings that you can play with to increase/decrease the amount of detail they provide.
P.P.S I haven't added any upscaling in this initial version but I will in the next. I'm considering SeedVR2 as the upscaler.
FAQ
Comments (28)
I have a character lora on Z-Image, but not on SDXL or Chroma. Can I still have good results with it by switching the models to ZIT instead?
I think you should be fine just adding the character lora to the ZIT model pipeline. You'll probably want to set the ZIT denoise setting to 0.70 just so you get as much of the character as possible without changing the rest of the image too much. There might be a little trial and error needed to find the right denoise value.
You'll also want to make sure your prompt accurately describes the character so that the Chroma generation can at least get pretty close.
It should work fine with any kind of Chroma and model loader. BF16, FP16, FP8 or GGUF shouldn't matter as long as it provides a model output.
Outstanding idea and implementation. Props to you.
Thanks!
Hi, I made this using a workflow modified from yours. Those modifications are mostly visual, but you may wanna check it out, just saying. 😘
https://civitai.com/images/114096031
Very cool, great job!
doesn't work, all text inputs are blocked except for detailer, detailers themselves have messed up settings with missing inputs. also could use a sfw version
Which version of the workflow are you having trouble with?
close Nodes 2.0
Took me a few goes experimenting with the Chroma sampler/scheduler to find the sweet spot, but once you find it, it's super sweet. Probably the best (most realistic) skin textures I've ever achieved. Thanks GBRX, another awesome addition to your collection. Cheers 😽
Awesome, thanks!
Hi friend. I'm having trouble getting my LoRa models to work. I have models in Z Image and SDXL. I don't know which LoRa I could put in the other LoRa node. Or how I can get it to use the LoRa from my model? Thanks
Which workflow are you using?
@GBRX Hi friend. GonzaLomo Z-Image Refiner Workflow
@iamaritomix545 Yes, but which version? CZXL, v1.1 or v1.0?
@GBRX gonzaLomo_CZXL1
@iamaritomix545 For that workflow it would only make sense to use the Z-Image lora. Just select it in the Power Lora Loader in the Z-Image Turbo Refiner group. The only problem is that there is a bug in the workflow. The Z-Image Power Lora Loader is not connected to the Refiner KSampler so you'll need to do that manually. I'll have an update for this workflow soon that fixes this.
@GBRX This is what I am doing with your workflow. for the Chroma Base, I have no character lora applied. For the zimage node, I am selecting my character lora and connected the lora to the refiner ksampler like you mentioned. It will not generate my character from the zimage lora. Thoughts? It seems to be similar problem that the pervious person was having.
@DirtyBeagles What is your denoise setting on the Z-Image refiner sampler? It probably needs to be at least 0.60, if not higher, for the Z-Image lora to be fully in effect. Idally, you would want to use the same dataset to create a Chroma lora as well and apply them both, if you have access to the dataset.
I like the refining approach! Z-Image is so great, but lacks in certain areas. Question: Your workflow uses bbox/Nipples.pt and you link in the description to where is supposedly found. However, there are no YOLO models with that exact name. When I download nipples_yolov8s.pt and use it in the detailer, it just isn't picked up. It has no effect. Any guidance here?
After updating ComfyUI and all the nodes, it is having an effect. However, I've gotten much better and more realistic results using bbox/female_breast-v4.2.pt. Just my feedback on your 1.1 workflow. You generate really high-quality images, so I'm sure I still have something off, but it's much better now!
@crankybuttercup Thanks for the feedback. I need to look into the issues with the bbox models and possibly host them myself.
@GBRX Thanks for sharing the detailers! I can confirm they give better results!
@crankybuttercup Great! I wish I had a better solution. Those detailers have been a pain.
Nice work, friend. Thanks for sharing it with us.