Illustrious--------------------
Same dataset as pony, but seem very different, much more closer to the style
Little bit over-trained, using around 0.8
Tested on waiNSFWIllustrious_v130.safetensors
DPM++ 2M SDE Karras
34 step
Hires.fix + Adetailer on the post images
don't really need Adetailer for this, some time Adetailer will smaller the eyes.
896 * 1152
POS: 1girl,
NEG: anime, 2d, bad quality, worst quality, worst detail, sketch, censor, blurry,
same dataset
new understanding
new model
new training ways
Video gen with wan2.1 :)
V2--------------------
For goodFitPonyv4.1 or realistic models. + lower size
Training on 500+ images from overwatch CGIs Sort image set from v1, delete most of the D.va image due to weird face. the overall face should be cuter than v1.
All faces are From Mercy, Tracer, Widowmaker, Mei, D.va, Kiriko, and Sombra. No male face has been trained. I will make a spread Lora for all CG styles of characters.
New training ways will cost this Lora much more cartoon, than v1 it will be more blurry, on anime or cartoon checkpoints.
-
All Post images are made on GoodFit-PonyV4.1 with
Steps: 35,
Sampler: Euler A SGMUniform,
CFG scale: 7,
Size: 896x1152,
Hires.fix no adetailer. (still needed when far)
v1---------------------
Training on 700+ images from overwatch CGIs. Aiming for the style of the overwatch CGI, not for a special character. Face should be all mixed in with the styles.
All train images are on 1024*1024
using settings other than 1024*1024 may give a funny result
highly recommend using ADetailer
The size has gotten up to close to 900MB, tried to lower the size but seems like it didn't hold the style well. This version should be pretty strong on different kinds of checkpoints.
Works great with the Pony itself on weight 0.9 or lower(pony without tune seems to understand it very well, but the result is just too blurry, I think you should be able to fix it with some extra prompts)
With basic prompts, the best res