[Genshin Impact] Linnea (Beta version)

Re-trained with 141 images.

Training parameters:

• Max train steps：2000

• Epochs：29

• Batches per epoch：141

• Batch size：1

• Gradient accumulation steps：2

• Effective batch size：2

• Learning rate：5e-4

• LR scheduler：cosine

• LR warmup steps：100

• LR cycles：1

• Optimizer：AdaMuonLoRA

• Max grad norm：1.0

• Mixed precision：bf16

• Loss type：l2 (MSE)

• Weighting scheme：uniform

• Min-SNR gamma：none

• Noise offset：none

• Multi-resolution noise：none

• Timestep sampling：sigmoid, scale 1.2

• Zero terminal SNR：no

• Gradient checkpointing：Yes

• Attention implementation：SDPA

• Latent caching：to disk

• Text-encoder output caching：Yes

• VAE：chunk size 64, 2-D mode, cache disabled

Network (LoRA) parameters

• Module：networks.lora_anima

• Network dim：64

• Network alpha：64

• Network dropout：none

• Rank dropout / module dropout：none

• Weight-norm scaling：none

• Train UNet only：Yes

• Train text encoder：No

• Train LLM adapter：No train_llm_adapter=False, llm_adapter_lr=0)

• Trainable parameters：91,750,400 (560 matrices)