Training Data

concepts: baseskin, blackchroma, bluechroma, greenchroma, pinkchroma, redchroma, tealchroma, whitechroma, yellowchroma

tags: armlet, bare shoulders, breasts, cleavage, detached collar, elbow gloves, eyes, facing away, frills, from behind, from side, full body, holding, long hair, looking at viewer, mermaid, mermaid tail, pumpkin hat ornament, solo, staff, transparent background, witchhat, witchdress

Trained on 45 images @ 45 epochs (2025 steps total) of in-game model 5 of each skin variant baseskin, blackchroma, bluechroma, greenchroma, pinkchroma, redchroma, tealchroma, whitechroma, yellowchroma

Training Data here: https://imgur.com/a/jBOVxDy

Weight Data Type: bfloat16

Override Unet Data Type: bfloat16

Override Text Encoder 1 Data Type: bfloat 16

Output Data Type: bfloat 16

Latent Caching: True

Clear Cache Before Training: True

Optimizer: AdamW

Beta1: 0.9

EPS: 1e-30

Stochastic Rounding: True

Beta2: 0.999

Weight Decay: 0.01

Fused Back Pass: True

Learning Rate Scheduler: Constant

Learning Rate: 0.0003

Learning Rate Warmup: 200

Learning Rate Cycles: 1

Epochs: 45

Batch Size: 1

Accumulation Steps: 1

Learning Rate Scheduler: None

Attention: XFORMERS

EMA: OFF

Gradient checkpointing: True

Train Datatype: bfloat16

Fallback Train Datatype: bfloat16

Autocast cache: true

Resolution: 1024

Train Text Encoder 1: True

Stop Training After 45 Epochs

Text Encoder 1 Learning Rate: 3e-06

Text Encoder Clip Skip: 2

Train UNet: True

Stop Training After 45 Epochs

Unet Learning Rate: 1e-05

LoRA Rank: 16

LoRA Alpha: 1.0

Dropout probability: 0.0

LoRA Weight Datatype: float32

The LoRA seems a little off x.x

The concept tags don't seem to properly work for the coloring & dress.

I might work on the tags later on... the hair & face comes out fine but the dress, gloves, hat, & staff doesn't most of the time for some reason...

I think the issue is the chromas were tagged the same witchdress & witchhat when I should've tagged them appropriately...

The pumpkin hat ornament tag has issues with placing multiple pumpkins in the image outside of the hat area.

Everything should be adjustable hair, breast size, dress, eye color, gloves, & hat as they were tagged.

I suggest removing the tag mermaidtail if you want to generate only halfbusts.

It's my second lora I've trained @_@ so if anyone has any tips I'd like them ❤️