Make sure to use UI that supports v-prediction, otherwise you will get a noisy blob!
See about version for credits!
Main parameters:
Steps: 28
CFG: 4
Sampler: euler a (euler for v6 and lower)
Scheduler: normal
Im also using CFGZeroStar and sometimes Self-Attention Guidance: default on difficult backgrounds)
Positive prompt:
number of characters (1girl/1boy), character name, artist prompt, scene/environment/camera angle, action, expression, items, quality prompts (masterpiece, best quality)some heavily stylized characters might require prompt: realistic
you can also try to put nose into faceDetailer
Negative prompt:
lowres, worst qualityurl against watermark
lineart, sketch, dot nose for realism
Enjoy..
Description
FAQ
Comments (2)
Thanks for releasing a new version after all! I have tried it briefly and I regret to say but I am still stuck with v3 due to much better prompt following for now.
I will check the 2DN base which was merged in v10 if this is causing the deterioration compared to v3 (I was trying to prompt for a group of people and some other non-solo setups which failed in v10 but worked in v3).
Anyway, I would appreciate further releases and hope you continue working on this merge @blinkdotleh!
I have looked into it, turned down 2DN weights to 20% (cause some pictures were too different at 33%) and played with the weights of my translucency-style and pure style LoRAs, also changed them into lower step versions for less chance of overfitting and damaging prompt that way. Making this model is a dance between translucency, darkness, style and prompt adherence, it's kinda hard, but fun. I'm still testing it, but v11 should be on the way soon.