The ultimate goal of this model is to provide an experience very similar to the already fairly competent base of NoobAI v-pred, while fixing up rough edges. Many other merges suffer from the bimodality of either having good prompt adherence (closer to base noob) or good default aesthetics (closer to illustrious).
Ideally, both can be encapsulated in a model without sacrificing too much model knowledge to achieve this.
Recommended Settings
For optimal results, we recommend the following inference parameters:
Sampler:
EulerorEuler AScheduler:
NormalorBetaSteps:
16-24CFG Scale:
3-6Resolution:
For general use:
832x1200(or similar aspect ratios with a total area around 1024x1024)For V9.1 and above: Can natively handle
1536xresolutions.
Description
Additional training to incorporate better style control, unfortunately I'm a dunce and botched training settings, but it still works
FAQ
Comments (4)
Great model! It has all the knowledge of the original NoobAI model, but at the same time it allows generating images in 1536px res. The only thing I’ve noticed is that at this resolution incorrect anatomy starts to appear, legs are too stretched out or arms, and this happens in almost every image. Other than that, everything’s great! As far as I understand, it’s also pretty difficult to change the style using artist styles or general styles like “anime screenshot.” At least I couldn’t manage it, maybe I was doing something wrong.
Strange; it's rather responsive to styles in my and others' testing; one thing that I have observed though is that masterpiece, best quality can be a little overbearing. 43k styles were explicitly trained at the front, which may or not impact anything, nevertheless at the very least I'd recommend dropping quality tags and putting the style tag of choice at the front of the prompt
edit: I believe the tag is anime screencap, not anime screenshot; seems to work for me anyway: https://files.catbox.moe/mhxpq8.png
velvet_toroyashi I tried adding styles at the beginning of the prompt, and it really works better. So, sorry for misleading you earlier.
GromForever All good :) I should add that to the usage instructions