Dataset: 121 images taken in-game with character creation screenshot tool

Model INZOIV1A summary:

- Base Model: Stable Diffusion v1-5

- LoRA Type: Standard, LyCORIS preset (full)

- Precision: fp16

- Optimizer: Adafactor, huber loss with SNR schedule

- Resolution: 768x768, bucketed between 256–2048

- Training:

- Epochs: 8, sampling every 4

- 121 images in dataset

- Batch Size: 3, gradient accumulation: 1

- Learning Rate: 0.0001 (UNet and text encoder)

- Adaptive Noise: Multires noise discount at 0.3; Original noise offset

- Output:

- Format: bf16 safetensors

- Sampling: Euler-A, prompts "inzoi style, a man" and "inzoi style, a woman"