This is the workflow I use the most. It's a flow I borrowed in part from another user, who I unfortunately don't remember, maybe @GBRX. I made some changes to adapt it to my work. This flow has helped me a lot in improving my work and I hope it can do the same for other users.
Description
FAQ
Comments (5)
I do not know what I am doing wrong. Both preview pics look great but the final image (i deleted the saveImagewithMetaData and replaced it with another Preview Image) looses all the details. The skin looks much better on the preview images. The final image looks like plastic and all skin details gone.
That's strange. It's normal for there to be a difference between the first and second previews because the second one has been refined, but that should already be more or less the final result. From there, only the face/eyes refinement is done... Did you happen to add LoRa in the refiner section?
However, without seeing your entire flow and the parameters you have, it's difficult to answer. Maybe try putting saveImagewithMetaData back in, create an image, and upload it. I can try to take a look at it.
It was a bit complicated because the node SaveImageWithMetaData is no longer compatible with current ComfyUI versions, but I found a fix.
Anyway, here is the end result:
https://civitai.com/images/122493971
Here is the image after the refiner:
https://civitai.com/images/122493972
I wanted to use the workflow with a different checkpoint. So I adjusted the cfg, sampler, and scheduler. But as I said, all the details will be ironed out in the end.
@tolstoi2003555 From what I can see, the parameters are correct for the model you are using. The only thing I would change is that I think 30 steps are too many for the Refiner. Furthermore, in my experience, it is better to use fast models that require fewer steps for the Refiner. You are basically generating at 30 steps and then refining at 30 steps, which is probably excessive. Try lowering the refinement steps, or working on 30 steps in total, then setting the first KSampler with a stop and the second one reopening from that stop, but for this you need to restructure the workflow quite a bit.
In the meantime, I would suggest trying to lower the refinement steps and see what comes out, then maybe try putting a faster model on the refiner and see how you like it better.