Jimwalrus
Well-Known Member
- Sep 15, 2021
- 1,102
- 4,197
I appreciate this isn't the Stable Diffusion prompt sharing and learning thread, but I've had a look at the prompts and parameters you used.
Here's some feedback, using that rather nice second image for examples:
Firstly, I'd recommend against using any celebrity pre-baked into Stable Diffusion. They're never any good if you want a true likeness and can often result in slightly wonky faces. In this instance it's not too bad and easily fixed - see later. Incidentally, good choice in using Irina - she has the face for 'Barbarian Queen"
Should you want a true likeness of a celeb, download a TI or LoRA from Civitai or elsewhere. Or train one yourself if there isn't one or it's not giving what you want.
Secondly, too many steps. 150 is far too many for Euler_a. Anything more than about 40 is probably into the realm of diminishing returns, 70-80 is the point at which anything more is largely a waste of time and risks introducing unwanted elements. Higher step numbers work better with some samplers, but Euler_a isn't one of them. Sometimes, less really is more!
Here are the same prompts etc, at 40, 60 & 80 steps instead:
Thirdly, if you have the facility, use Hires.fix within the initial generation process, not upscaling afterwards. Even if it's only to increase it by 8 pixels in each direction, it allows considerable improvement of the image. A far better use of your GPU (or Google Colabs tickets) than excessive generation steps.
Here is 50 gen steps, 50 Hires steps, ESRGAN_4x, 0.33 denoising strength, to just x1.05 upscale (so no real effort for the GPU):
Fourthly:
If you're going for anything like photorealism, try using Restore Faces. Doesn't always work, but usually does. IF the style is more cartoony then it screws it up, but for an image like this, it should help:
Fifthly:
Try using a few negative prompts. They can reinforce concepts as well as guard against things going wrong.
You can also download pre-baked negative embeddings from Civitai and other places.
Adding 'Unspeakable-Horrors-32v', 'bad-hands-5' and 'easynegative' makes for a marked stylistic change. Not 100% sure I like it in this instance, but it's worth bearing in mind. Also, there wasn't really anything wrong with the original image that needed fixing with negative prompts.
Finally, if you can, try changing the aspect ratio of the image. Keep one side as 512 pixels, then adjust the other.
You're going for a standing person, so best to have a taller image.
Here's 50 steps, 512x832, with 50 steps of x1.05 upscaling, no negative prompts:
Here's the final result, upscaled within the initial generation process by 3x to 1536x2496:
(Apologies everyone, I know it's a big file to download, but I wanted it to display in full image size)
This resolution is about the limit of my GPU without using Tiling.
