Fervent Tempo ∆
Character: Rozaliya Olenyeva (Honkai Impact 3)
Battlesuit: Fervent Tempo ∆, original outfit
Pixiv post here
Usage (Important!)
Compatibility
Should work with most anime-like checkpoints. Tested with AOM3, BOM, Hassaku, Grapefruit, PVC-AOAOKO, Cornflower X, etc.
Weight
Recommended 0.825; 0.75~0.90 range works nicely depending on how much flexibility you need.
Trigger Words
No trigger words are required. However:
You can put (mismatched stockings, long hair:1.15) on prompt and (tail, mechanical tail, weapon, sword, holding, holding weapon, holding sword:1.15) on negative prompt for better consistency.
Other
Image Details
Preview images use AOM3 or BOM, unless stated otherwise.
VAE used is orangemix (nai)
Training Details
May vary based on version. See individual versions for details.
Description
This one took me a while until I was satisfied enough to post here, I've been experimenting and stuff.
Improvements:
File size 4 times smaller than Release 1!! (dim = 128 -> 32)
Better at retaining artstyle of the base checkpoints used to generate ∆.
More consistent outfit details. Notably, conveys the translucency of her jacket better, and the "power button" mark on the torso of her jacket.
Limitations:
Stockings still may be plain black instead of blue + pink. Adding (mismatched stockings:1.15) to your positive prompt mitigates this.
Tails may appear uncontrollably. Adding (tail, mechanical tail:1.1) to negative prompt may help, but they aren't captioned in the training data. Maybe I should have so they could be separated. Will probably be fixed in Release 3.
Her weapon may appear uncontrollably. This is tagged though, so adding (weapon, sword, holding, holding weapon, holding sword:1.1) to negative prompt mitigates this, but it's not perfect.
From the viewer's perspective, the right side of her jacket may not be black, and instead gets colored white at times. This is gacha and not controllable.
Some limitations with H content due to training dataset all being SFW.
Training Details (w/ bmaltais/kohya_ss; see Training Data File for full LoRA training parameters extracted with kohya-ss/sd-webui-additional-networks)
Trained on: nai.ckpt (925997e9)
Training Batch Size: 2
LR: 0.0007, UNet LR: 0.0007, TE LR: 0.000175, Scheduler: Cosine w/ Restarts
Epochs: 3, Dim: 32, Alpha: 16, Resolution: 512x512, Clip Skip: 2
*EasyNegative in the example images is an embedding from here.