After fighting for a while, i feel this is good enough for a release! AnBan got a Vpred version!
TL;DR
CFG: 5.5
sampler: Euler (automatic)
steps: ~30
CLIP skip: 2 (you can do without if you want, but it works with it)
Positive prompt:
IllusP0s, <insert here your usual prompt>
Negative prompt:
IllusN3g,signature,username,logo,bad hands,mutated hands,fused fingers
Long story
The story so far is covered here:
Basically, after the first article, i had AnBanV_V0 and with the second article, i had AnBanV_V1alpha.
After some testing, i detected some artifacts with the V1alpha and the colors were a bit too strong. And hands got messy a bit... That's why i tested a new merge with V0 to get the V1beta:
My AnBan LoRA, Epoch 2: 0.38 (0.4 felt a just bit too much, i ballparked it to this number)
Unified Theory LoRA: 0.12 (a "cure-all" LoRA training on high esthetic picture, number choosen out of a hat, i wanted something between 0.1 and 0.2)
Detailed Hands: 0.1
But to add just a touch of varnish on it, i wanted to add a smidge of anime look back on top of it. That's when i got the update about the release of Galena CAT for Illustrious. This is a checkpoint, not a LoRA, but CitronLegacy was kind enough to explain that it was a merge of a trained LoRA on top of WAI-NSFW-illustrious V7.
So, i applied a "Add difference" merge at low multiplier (0.1) to get just a very small touch of this goodness in my model: AnBanVpred_V1= AnBanVpred_V1beta + 0.1 x (Galena - WAI)
And so, here is all the steps in one matrix, from the original NAI-XL to the final AnBanVpred:
I have added manually the metadata in the model to try and make sure it is detected as a V Prediction model:
{
"modelspec.predict_key": "v",
"modelspec.title": "AnBanVpred_V1 1",
"modelspec.sai_model_spec": "1.0.0",
"modelspec.architecture": "stable-diffusion-xl-v1-base",
"modelspec.implementation": "sgm",
"format": "pt"
}
And forge seems to be playing nice and detect the model as expected, even if i don't set the noise schedule to "Zero Terminal SNR" ^^
And to finish this article, i'll add again: this is a V prediction model. A1111 will NOT work with it and the usually recommended sampler is Euler (not Euler a, not DPM++ 2M or some other fancy sampler), as demonstrated here (PS: no Adetailer, 30 steps), but feel free to experiment:
A1111, DPM++ 2M (Karras)
Forge, DPM++ 2M (Karras)
Forge, Euler (Automatic)
NB: Using FreeU and Self-Attention Guidance (both integrated in Forge) is possible and tend to give more contrasted and colorful results with more details. That's a personal choice to make :D