This LoRA model is in testing stage, it should not be anticipated to get highly amazing results.This is my first LoRA.<h2 id="notes:">Notes:</h2><ul><li>The dataset contains few features related to Lumi, thus this model may behaves not well in generating images related to it.</li><li>Entire dataset was tagged with only natural language and I did not test if tag-string prompt works well.</li><li>If there is any error, please let me know.</li></ul>

[Genshin Impact] Linnea (Beta version)

Re-trained with 141 images.Training parameters:<ul><li>Max train steps：2000</li><li>Epochs：29</li><li>Batches per epoch：141</li><li>Batch size：1</li><li>Gradient accumulation steps：2</li><li>Effective batch size：2</li><li>Learning rate：5e-4</li><li>LR scheduler：cosine</li><li>LR warmup steps：100</li><li>LR cycles：1</li><li>Optimizer：AdaMuonLoRA</li><li>Max grad norm：1.0</li><li>Mixed precision：bf16</li><li>Loss type：l2 (MSE)</li><li>Weighting scheme：uniform</li><li>Min-SNR gamma：none</li><li>Noise offset：none</li><li>Multi-resolution noise：none</li><li>Timestep sampling：sigmoid, scale 1.2</li><li>Zero terminal SNR：no</li><li>Gradient checkpointing：Yes</li><li>Attention implementation：SDPA</li><li>Latent caching：to disk</li><li>Text-encoder output caching：Yes</li><li>VAE：chunk size 64, 2-D mode, cache disabled</li></ul>Network (LoRA) parameters<ul><li>Module：networks.lora_anima</li><li>Network dim：64</li><li>Network alpha：64</li><li>Network dropout：none</li><li>Rank dropout / module dropout：none</li><li>Weight-norm scaling：none</li><li>Train UNet only：Yes</li><li>Train text encoder：No</li><li>Train LLM adapter：No train_llm_adapter=False, llm_adapter_lr=0)</li><li>Trainable parameters：91,750,400 (560 matrices)</li></ul>

Files with SHA256: a3d1d611e3855e80b9a1958dd2c999766216d742d6a7db7ae3a2f0e3420d241d