ARRISA (AI LIMIT)版本v1.0 (ID: 1023169)

Model trained on images of Arrisa from upcoming souls-like AI LIMIT

描述:

Dataset: 32 images taken in-game with external screenshot tool

Model ARRISA summary:

- Base Model: sdxl_base_v0-9

- Precision: fp16

- Learning Rate: 0.0003 (UNet), 6e-05 (text encoder)

- Epochs: 12, steps per epoch: 80, total steps: 960

- Mixed Precision: fp16 with gradient checkpointing

- Optimizer: AdamW 8-bit (weight_decay=0.1, betas=[0.9, 0.99])

- Scheduler: Cosine with restarts, warmup steps: 48

- Gradient Accumulation: 1, Max Grad Norm: 1.0

- Token Length: 225, Prior Loss Weight: 1.0

- Noise and SNR:

- Multires Noise Discount: 0.3, iterations: 6

- Min SNR Gamma: 8.0

- Network: LoRA (module: networks.lora, alpha: 4, dimension: 8)

- Seed: 42, Cache Latents: True, Low-RAM Mode: True

- VAE: stabilityai/sdxl-vae

- Output Name: ARRISA, Session ID: 1055024336

- Source SD Model: hollowstrawberry/67AB2F

训练词语: left arm active,left arm inactive,Arrisa