Type | |
Stats | 171 1,336 |
Reviews | (29) |
Published | Oct 8, 2023 |
Base Model | |
Training | Steps: 3,710 Epochs: 35 |
Usage Tips | Clip Skip: 2 |
Trigger Words | UmeIttla, colored skin, orange eyes, oni horns, skin-covered horns, short hair toned abs |
Hash | AutoV2 BAD3FB48BA |
I have opened up a request form
The Oni OC Ume from the artist Ittla. Works best between 0.8-1.0. Training done on NAI.
Trigger V3: "UmeIttla, colored skin, orange eyes, slit pupils, oni horns, skin-covered horns, short hair" always needed, add "toned" and "abs" for more accurate physique.
Suggestions/notes:
V3:
Added "skin-covered horns, short hair" to training which fixed hair turning into horns during hires.
Trained slightly longer than V2 which helped improve color and consistency.
Flexibility on clothing is still good, check samples from outfits ideas.
V2:
Trigger: "UmeIttla, colored skin, orange eyes, slit pupils, oni horns"
Extra tags were added in training on this version as they were the less stable parts of the character, now those features are more consistent across images.
"Slit pupils" are still hit/miss depending on model/hires settings so I suggest higher hires denoise as it can help create cleaner details.
Very flexible clothing wise as there is no set outfit.
Check above samples for outfits ideas (Some are based on source images)
Warning: Works best on animated style models. To use consistently on other less anime models you need to add "Purple skin, black hair" to avoid skin going pale.
A shot at remaking my first model on civitai as a LoCon with new training methods I have learned. It came out as nice upgrade with the helper tags being added to training as that way they stay a lot more consistent at the higher training resolution. I hope you all enjoy.
Feedback and reviews are always appreciated.
Nerdy training numbers (V3):
Trained on D8Dreambooth trainer
Optimizer: AdawW Dadaptation
Training resolution: 768
Unet LR: 1
Tnec LR: 1
Unet weight decay: 0.016
Tenc weight decay: 0.02
35 Epochs - 3710 Steps
Trained on 53 images using Reg images.