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
v1.0 of this workflow used SDXL to refine the whole image which turned out to be less than ideal. With v1.1, I'm taking a different approach that I'm not entirely sure is necessary, although it does seem to give me the best result.
Essentially, I'm taking a similar high/low noise approach to the way WAN 2.2 video generation works. 3 steps on "high", return with noise, and then 7 steps on "low". In my A/B testing this seemed to give a better result than just running 10 steps in one sampler. I'm also generating at a very high resolution 1536x1920 and not doing any upscaling.
Theoretically, the "high" noise sampler could have different sampler/scheduler/CFG settings and possibly increase image composition and quality further.
FAQ
Comments (10)
I prefer the simplicity of 1.0 and feel like I have more control over the output. 1.1 isn’t bad, I’m just not sure the results are worth the extra steps. Appreciate the experimentation… 1.0 is the best ZiT workflow out there. The combination of nsfw fix with 4-step DMD, the detailers, and upscale make for 10/10 images every time. It’s super fast too! Looking forward to future explorations on ZIT workflows from OP.
Awesome, I really appreciate that feedback. I was a bit torn between the two versions. 1.1 is slower but the result is more pure ZIT so I like it for that reason. I'll continue offering both types of ZIT workflows with future versions since they both have value.
@GBRX glad to hear it! one feature request: make it so enabling detailers + upscale w/o needing the refiner model works. right now it seems the refiner model is required for any of the downstream nodes to have any effect on the output. :)
@cleonthethird The SDXL refiner model is necessary for the detailers to work properly. Are you suggesting that the detailers use the ZIT model to do the refining?
@GBRX Ahh I didn't realize SDXL was needed for the refiners. I was hoping to be use the refiners with Z-image turbo only during the initial image generation, ideally to save a bit of time if the SDXL de-noise is not desired but the adetailers are.. Currently I just turn the SDXL strength to .1 or lower to minimize any intricate detail loss, but it seemed like an extra step that I was hoping to be able to eliminate to save time.
@GBRX btw I find myself mostly disabling the refiner stage (unless needed for NSFW improvement)... I just realized the output is being stored in temp and not saved in the output dir. I did some research and found it's due to the image preview node and not save image. I was going to try to edit the workflow later (I'm a Comfy n00b) to create a variation that removes the SDXL refiner steps (maybe trying to add upscale only). It would be nice if the refiner stage was turned off if the images were saved in output (and not temp). Not sure if this is helpful, but just sharing my experience! Thanks again! Cheers.
@cleonthethird I'm going to unify v1.0 and v1.1 when I do v2.0. It should give you more options for how/when the refiners are applied. Remember, the refiners are only really needed for nudity. Once we get better nudity baked into the models then the refiners won't be necessary.
Beautiful Work, and a huge volume of it, have you found a way to make money with it as I'm sure this takes up a good deal of you time and you're not cranking this out on a RTX 3060 12 GB Vram, you probably own a rack of servers.
Ha, thanks! I've got a RTX 3090 and RTX 4060 that I do everything on. I could use a rack of servers though!
killer