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"
