Experimental

Dataset: 112 images, captioned with Qwen3-VL-8B-Instruct-abliterated+Trigger Word "double penetration"

Training Tool: Ostris AI-Toolkit

Training Model: Tongyi-MAI/Z-Image-Turbo

Settings:

• Quantization: None

• Data Type: BF16

• Batch Size: 1

• Gradient Accumulation: 1

• Steps: 24000

• Optimizer: AdamW8Bit

• Timestep Type: Weighted

• Timestep Bias: Balanced

• Loss Type: Mean Squared Error

• Linear Rank: 32

• Learning Rate: 0.0001

• Weight Decay: 0.0001

• Resolution: 512