This is my long running workflow cleaned up and streamlined for y'all so you get good results using my models out the gate.
There are plenty of notes in the workflow with explanation. I recommend using my SDXL models: https://civarchive.com/models/605283/donutsdeliverymix
Quick Tip: The prompt in Face related section automatically gets included with the main prompt and does not need to go both places.
Hotfix 9 Jan: location for sam model in instructions was wrong and has been changed to /sams/
Help and Questions
Got any questions you can catch me in my discord server: https://discord.gg/kt2hCJHuKp
Description
Added individual face+eye detailer for male and female with gender detection (should work out the gate)
Added IPadapter
Added module on/offs (Does not currently work for upscaling modules)
Added logo (loaded from web)
Added hands detailer
FAQ
Comments (12)
I really like your model and wanted to use your ComfyUI workflow (BTW, appreciate your work!). However, when loading the workflow, I get a warning message with a huge list of missing node types. This might sound like a very dumb question, but since I am not that familiar with ComfyUI yet (coming from WebUI), I wanted to ask if you were able to give some hints how to get this working? Thanks in advance!
Thank you very much for the experimentally selected parameters in the primary generation and upscaling chain. I will only note that it is better to replace the upscaling model in the second and third blocks with others, since NickelbackFS after the first upscaling begins to strongly distort the internal lines of the portrait object.
And could you explain your note between the first two KSamplers:
"This is so you can customize which loras are enabled at different stages"
which upscaling model do you recommend instead?
The first pass is done by two different ksamplers with different seed, this is so you can for example use different rescale CFG at different steps 0-20/20-40. You can also choose to bypass the lora for the last 20 steps this way which you might want if you want the lora only to apply to the look and feel instead of the composition/structure of the image.
@DonutsDelivery Thanks again, I missed it in another note about the LoRA Loader (Block Weight).
As for upscalers, it would be easier to connect 4x-UltraSharp everywhere, but then the effect of creativity in the 1st and 2nd stages of upscaling is weakened. Therefore, for now, I left the one you chose at the first stage, 4x_BS_DevianceMIP_82000_G at the second stage (it is already better, but this is not the final choice yet), and 4x-UltraSharp at the third stage.
And one more piece of advice, if you allow me: if you are not going to use ControlNet in this workflow, then put the Kohya Deep Shrink node after the LoRA loader - even with the default parameters, it shifts the aesthetics from illustrative to more natural-realistic.
Also, after the first paired KSamplers with the standard fixed VAE model (VAE Decode), I use TAESDXL in Anything Everywhere3.
And I completely forgot to mention that after the third upscaler and in the last block (Hands Detailer) it is better to use the Purge VRAM node from the LayerStyle extension.
@windlike_gust Thanks for the tips. I'll play around with it and do some A/B testing to see if it's something worth putting in the new version. I primarily use nickelback because it preserved wet/oily skin instead of flatting it out like other models.
that purge vram node could have saved from a lot of trouble lately...
@DonutsDelivery There is one significant problem in the implementation of this absolutely wonderful method, which extends the capabilities of ComfyUI: BlenderNeko has not updated its pack for a long time, despite numerous requests and error messages. In all available examples (including yours), only dmpp_2m works in Unsampler, which is practically useless.
Among the error comments from other GitHub users, I saw this one:
"i had the same problem, you can replace the unsampler from this extensions with a combination of nodes from sample/custom_sampling, mostly SplitSigmas and flipSigmas
Edit: you an use the advanced latent operations for that, but getSigma still remains a mystery for me."
It turns out that it is necessary to use other nodes to perform this task.
@DonutsDelivery A little more: not all models contain VAE. Also, different node categories with different samplers process tone and color differently. So there is a situation when DetailerForEachDebug merges processed SEGS and even to the eye the tonal and color difference of the superimposed areas is noticeable. To compensate for such discrepancies, you need to enter your own VAE for each stage: for example, the standard one for linking the first two samplers, TAESDXL for upscales and xlVAEC_c9 (or c1) for DetailerForEachDebug.
In the last two scaling blocks, I liked how the 4xFFHQDAT model worked, but it slowed down the result considerably.
And despite the stepwise increase in SEED in the three-stage scaling, you also need to be careful that there is no significant gap in the initial values in the last stages, since this is very likely to cause severe visual distortions.
@windlike_gust Got very lousy results with TAESDXL, very blurred results. Deviance upscaler model is trained for anime, and ultrasharp is like a universal model. nickelback is superior at preserving skindetail but feel free to use any other model for your usecase. You can use the same seed when using 2 different ksamplers in the first pass, it gives very poor results. I've also A/B tested same seeds for upscaling but different seeds give more natural looking results. If your model does not have a VAE you gotta use a vae loader and plug it into anything everywhere node. deep shrink is interesting but it's usecase is generating higher resolution images with the same structual integrity as standard resolutions. It's not something I would consider adding to the workflow. Testing it out I noticed some A/B tests gave good results but also less creative results. I'm gonna continue to use BlenderNeko unsampler as the A/B tests came out better than not having it. If there is a better alternative I need a guide how to do it as i dont even know what a sigma is.
@DonutsDelivery You were right about the upscaler model. As for saturation degradation compensation, I used the KJNodes Color Match. It is indeed wrong of me to compensate such thing with VAE.
And again, many thanks for all the comments and clarifications!
Sorry to bother you about this but whenever I try to use your workflow the image doesnt seem to pass through the whole workflow it stops after the second ksampler.. am I missing something? (I have used the manager to install all the missing nodes)
after the second ksampler... did you set the upscale models? or you can download the one i put i recommend, nickelbackFS