DJZ AssassinKahb [Cascade Test Lora]

V1 evaluation lora used to create a baseline.
Training settings were refined for the next versions;
V2 - ADAFACTOR, Rank 16, Alpha 1
V3 - Rank 32, Alpha 1
v3b - rank 32, alpha 8
v3c - rank 32, alpha 4
V4 - were all trained with ADAFACTOR using revised settings. Rank 64, Alpha 1

Round 2

V5 - Trained with ADAFACTOR with final settings, Rank 32, Alpha 1, reduced learning rate, LR 2e3

Switched to PRODIGY

V6-early = ADAFACTOR, Rank 32, Alpha 1
V6 = ADAFACTOR, Rank 32, Alpha 1

offset noise 0.06
V7-early = PRODIGY, Rank 64, Alpha 1
V7 = PRODIGY, Rank 64, Alpha 1

initialD 1e-06

V8-early = PRODIGY, Rank 64, Alpha 1
V8-mid = PRODIGY, Rank 64, Alpha 1

V8 = PRODIGY, Rank 64, Alpha 1

as shown in the video, early Epochs seem to perform better beyond txt2img.

Study presentation: https://www.figma.com/file/NM9dSIwcKoyZpOSYqi4pKs/OneTrainer-Stable-Cascade-(LR-adjustment)

Models: https://civitai.com/models/320332 Dataset: https://github.com/MushroomFleet/assassinKahb-1024

Article: Training Stable Cascade with OneTrainer

OneTrainer: https://github.com/Nerogar/OneTrainer

my preset for OneTrainer: https://pastebin.com/tLri1HdU

.json file, place inside OneTrainer\training_presets

which are trained to compare the Rank setting.

test prompt:

• AssassinKahb style a demonic looking skeleton holding a sword with red hair

txt2img-lora: workflow

img2img-lora: workflow

txt2img-vision-lora: workflow

txt2img-remix-lora: workflow

At the time of posting the Lora implementation is incomplete in most webUI's You will be able to use this if you have updated your ComfyUI to the latest version.