<h3 id="model-training-summary:-8s9vsu560">Model Training Summary:</h3><ul><li>Dataset: Continuing to gather for next round</li><li>Captioning: Generated using Koyha_SS with GIT captioning.</li><li>Training Platform: Civit’s onsite trainer, targeting 1,050 steps for this round.</li></ul><h3 id="training-configuration:-hlieoyv5o">Training Configuration:</h3><ul><li>Engine: Kohya</li><li>Learning Rates:<ul><li>UNet LR: 0.0005</li><li>Text Encoder LR: 0 (no adjustment)</li></ul></li><li>Clip Skip: 1</li><li>LoRA Configuration:<ul><li>Type: LoRA</li><li>Network Dimensions: 2</li><li>Alpha: 16</li></ul></li><li>Training Details:<ul><li>Repeats: 30</li><li>Resolution: 1024</li><li>Batch Size: 4</li><li>Max Epochs: 5</li><li>LR Scheduler: Cosine with Restarts (3 cycles)</li><li>Min SNR Gamma: 5 &lt;---- I find this important</li><li>Noise Offset: 0.1 </li><li>Bucket Enabled: True</li><li>Optimizer: AdamW8Bit</li></ul></li><li>Augmentation: No flip augmentation; captions were not shuffled.</li></ul><h3 id="notes:-r75hctwrt">Notes:</h3><ul><li>Resolution 1024, allowing for higher fidelity image training.</li><li>Repeats 30, potentially boosting the model’s understanding of the dataset.</li><li>A shorter target step count of 1,050 is being used for this version</li></ul><h3 id="next-steps:-d8wih9ttq">Next Steps:</h3><ul><li>Version 2 in progress gathering dataset/editing/updating</li></ul>

B3LLA_Flux_

First Version in Flux, seems to be decent." Recommend Flux Dev f8

Files with SHA256: d69b220be375f890a4793606aa50a97caf1b7bc3484de7c68e44b5a8bcad1f4e