Model Training Summary:

• Dataset: Continuing to gather for next round

• Captioning: Generated using Koyha_SS with GIT captioning.

• Training Platform: Civit’s onsite trainer, targeting 1,050 steps for this round.

Training Configuration:

• Engine: Kohya

• Learning Rates:

  • UNet LR: 0.0005

  • Text Encoder LR: 0 (no adjustment)

• Clip Skip: 1

• LoRA Configuration:

  • Type: LoRA

  • Network Dimensions: 2

  • Alpha: 16

• Training Details:

  • Repeats: 30

  • Resolution: 1024

  • Batch Size: 4

  • Max Epochs: 5

  • LR Scheduler: Cosine with Restarts (3 cycles)

  • Min SNR Gamma: 5 <---- I find this important

  • Noise Offset: 0.1

  • Bucket Enabled: True

  • Optimizer: AdamW8Bit

• Augmentation: No flip augmentation; captions were not shuffled.

Notes:

• Resolution 1024, allowing for higher fidelity image training.

• Repeats 30, potentially boosting the model’s understanding of the dataset.

• A shorter target step count of 1,050 is being used for this version

Next Steps:

• Version 2 in progress gathering dataset/editing/updating