Best settings:
Sampling Method: Euler A or Euler A SGMUniform - you can mess around with this
Sampling Steps: 20 ~ 60 depending on your sampling method
Hires. fix: Enabled
Hires Steps: 15 ~ 30
Upscaler: 4x_NMKD-YandereNeoXL_200k - You can test others.
Denoise Strength: 0.4
Upscale by: Preferably x2
CFG Scale: 6 ~ 9.5
Clip Skip: 2
Important:
Positive prompt must contain: score_9,score_8_up,score_7_up,
You can add "lineart, [artist tag], outline" for a more stylized result.
Negative prompt can be empty or contain the default pony negatives.
The promting format from Pony Diffusion V6 works well:
score_9, score_8_up, score_7_up, just describe what you want, tag1, tag2If you haven't already, I highly reccomend reading through the Pony Diffusion V6 XL model description.
I use Forge as my main UI because it has so many optimizations for running XL/Pony models on lower end GPU's despite me having an RTX 4090, generations are alot smoother. I'm not sure how this model performs on other UI's.
I haven't done much testing. If there is anything amiss, please let me know. Feedback is always appreciated :).
Post results and, most importantly, have fun!
Description
FAQ
Comments (4)
From my testing, the model looks pretty good, especially with prompts borrowed from your other model, EbaraMFCGPonyMix.
Great work.
Thanks!, yeah its a mixture of that model + another one I trained but is not ready for realease.
I tried a lot models based on PonyV6 on civitai, this model has a awsome anime body style, almost perfect to me. The only fly in the ointment is that the "knowledge" about the characters isn't very extensive
Thanks for the feedback!, There isn't many artwoks of actual characters in the original dataset so that could be the issue, I'll have to find a way to fix that without messing with the style of the model too much. You could try to use different character Lora's to generate the characters you want for now.