MilkyChicken Mix
This model is an attempt at introducing simple stylistic elements from my favourite models without losing the flexibility of robust prompting when desirable.
I am interested to see how it performs for you =]
Suggested settings:
CLIP Skip: 1 or 2
Size: 512x512, 496x768, 512x768, 512x900, 768x512
ENSD: 31337
Highres fix: Lanczos @ denoise 0.4 or Latent @ denoise 0.55
Sampler: DPM++ SDE Karras
Steps: 20 - 30
CFG: 7
VAE: none (baked)
Prompt:
(masterpiece), (best quality), (detailed), (realistic), 8k, (unnecessary in v2.0)
Neg:
cartoon, painting, illustration, (worst quality, low quality, normal quality),
Tips:
For better quality, try using more/less weight, i.e. (best quality:1.3), etc...For complex scenes, try using more steps in sampler, 25-50.
For better lighting, try describing it, i.e. (dramatic lighting), (sunshine), etc...
Description
V1.1
Same ingredients, new recipe!
possible output improvement
potential realism increase
less reliance on quantifiers, i.e. "best quality, masterpiece, realistic, 8k" etc...
obvious stolen prompts in showcase images (thank you if I stole yours!)
FAQ
Comments (3)
I've tried few generation and I must say that it's very clean, every generation didn't need much intervention. Most of them were already "ready to go". Good Job :)
That's great to hear! I'm glad that it is working well for you too. Thank you for the feedback =]
I say more:
I learned to distrust the various checkpoints lately out on CivitAI: expecially when the demo gallery is full of strange addings in the positive prompt and full of negative embedding in the negative prompt. Pimped like this, They will give good images but they wont follow the prompt you requested!