- Jul 6, 2021
- 414
- 6,410
- Jul 6, 2021
- 414
- 6,410
- Jul 6, 2021
- 414
- 6,410
- Jul 6, 2021
- 414
- 6,410
- Jan 30, 2024
- 133
- 1,756
- Jan 30, 2024
- 133
- 1,756
- Apr 15, 2020
- 241
- 1,352
I have Jenny & Roxy created (i made them over a year ago) but i purposely made them cartooney, but i think i'm going to go the way i did with Judith, and make their faces a little bit more realistic.
This is what they look like at the moment (cartooney face)
This is what they look like at the moment (cartooney face)
May I ask what your prompt and negative prompts are when you are crafting those beautiful tits? Your material is amazing!
- Jul 6, 2021
- 414
- 6,410
"ugly, deformed, bad quality, amateur drawing, beginner drawing, deformed, bad anatomy, deformed, white background, deformed hands, deformed feet, ugly eyes, ugly tits, watermark, 6 fingers" is what i use for a baseline but i found that the more steps i use the higher consistent quality i see along with a high CFG scale
- Apr 23, 2017
- 152
- 295
Thank you. When you say more steps, could you provide an example, ie the Goth girl one?
- Jul 6, 2021
- 414
- 6,410
This is what i mean, for that goth girl pic the steps were set to 50 and the CFG to 4.5
This is the result (with added inverted nipples)
Last edited:
Thank you for that. I completely overlooked that advanced option at the bottom. I am seeing some minor improvement, but nothing matching your quality. Do you also adjust the LoRA slider at all? I've kept it 1.0 but wasn't sure if I should decrease or increase it. Also, I'm including a prompt description of one of my attempts. For your pictures, are you doing a full on paragraph of description you might see in a novel?
- Jul 6, 2021
- 414
- 6,410
I usually keep the slider around 70 only going up to about 80/85 or down to 60/65
This is the exact prompt i used along with the pony diffusion model which resulted in this image
If you're still not getting something along these lines of quality just keep at it, repetition is really the best method i think
To get good quality, nothing will beat increasing the resolution and regenerating portions of the image.
That's why having a local installation is so important, since you can't do image to image in civitai.
Tile resample controlnet worked really well in 1.5 at keeping the shape when regenerating in image to image. You could set the denoise at 1 and be fine. But tile resample isn't available with sdxl. You can try other controlnets, but you'll have to set a low denoise.
I could train a 1.5 version of the LoRA so we can use tile resample. I'll have to think about it. I think with bucketing selected I don't have to rescale all of my training set images. But it would be better for the quality to crop the images rather than rescale, and that's quite a lot of work. The set has 700 images I think.
But considering that the style doesn't have a lot of detail, using tile resample isn't excessively useful. A simple image to image at low denoise on an upscaled image is pretty much good enough.
That's why having a local installation is so important, since you can't do image to image in civitai.
Tile resample controlnet worked really well in 1.5 at keeping the shape when regenerating in image to image. You could set the denoise at 1 and be fine. But tile resample isn't available with sdxl. You can try other controlnets, but you'll have to set a low denoise.
I could train a 1.5 version of the LoRA so we can use tile resample. I'll have to think about it. I think with bucketing selected I don't have to rescale all of my training set images. But it would be better for the quality to crop the images rather than rescale, and that's quite a lot of work. The set has 700 images I think.
But considering that the style doesn't have a lot of detail, using tile resample isn't excessively useful. A simple image to image at low denoise on an upscaled image is pretty much good enough.
Last edited:
- Jul 6, 2021
- 414
- 6,410
- Jul 6, 2021
- 414
- 6,410
That worked! Thank you! Going to make pregnant versions of the in game characters with this!
- Jul 6, 2021
- 414
- 6,410
- Jun 24, 2022
- 2,870
- 22,899
