Fan Art Dating My Daughter: Fan Art

5.00 star(s) 2 Votes

ccd17

Newbie
Nov 24, 2017
27
66
Right, I didn't pay attention to the low denoise, that's why it still looks like 3D and not more photorealistic. did you try with like 0.75 denoise or more ?
yeah I tried higher denoise, but the higher you go the more differences you get and the more detailed the prompt has to be to keep from having more changes introduced. I did upgrade to V7 and used restart, though bumped up to 75 steps as 66 was a not as crisp. The result was nicer but still somewhere between photo and 3d. I'm running through a bulk run of day 1 images to see how it works. Looks ok but it it seems to really want to give her a nose ring.
I agree it still has a bit of the original 3d look, but I really want to get to a point where I can bulk run with as little input as possible so higher denoise is not a great option. I am curious what would happen if I were to run an image through multiple times at a lower setting and see if it got more photoreal each pass.
 

Kryptozoid

Well-Known Member
Sep 3, 2016
1,201
2,500
yeah I tried higher denoise, but the higher you go the more differences you get and the more detailed the prompt has to be to keep from having more changes introduced. I did upgrade to V7 and used restart, though bumped up to 75 steps as 66 was a not as crisp. The result was nicer but still somewhere between photo and 3d. I'm running through a bulk run of day 1 images to see how it works. Looks ok but it it seems to really want to give her a nose ring.
I agree it still has a bit of the original 3d look, but I really want to get to a point where I can bulk run with as little input as possible so higher denoise is not a great option. I am curious what would happen if I were to run an image through multiple times at a lower setting and see if it got more photoreal each pass.
I think there's a way to do that automatically with Controlnet but I've never tried. I also think the quest to be able to batch process all the images of a game or even just 20 and expecting a consistent result is a bit futile right now. But kudos to you if you manage it.
In any case , controlnet like Openpose, Depth and Canny help to keep the image consistent while using higher denoise.
I did a bunch of tests just now and had good results (not sure if I'll post because I just did the start screen background image like 40 times with different parameters).
What could also help is to train your own embedding or Lora of Dee.
But then you will get into trouble whenever it's not only her in the pictures. So batch anything won't work in these cases and you'll be forced to use inpainting for faces or whole characters.
Not to mention that apart from all this, even background will change (slightly or a lot) from picture to picture, same for clothes. Even if you can limit both by using the same seed, it probably won't be enough to make it consistent.
 
5.00 star(s) 2 Votes