quarterturn/qwen-image-20b-ruri-rocks

A Qwen-image 20b LoRA for emulating the Ruri Rocks style and characters.

Note:

• Every caption in the dataset contains "Ruri no Houseki" so if you use that, you should trigger the LoRA
• The main characters were identified in the dataset images
• the dataset was captioned by Gemma3-27b via a custom script

Trained using musubi-trainer with the defaults for qwen-image using 796 1920x1080 images for 16 epochs using the following command:

accelerate launch --num_cpu_threads_per_process 1 --mixed_precision bf16 src/musubi_tuner/qwen_image_train_network.py     --dit /home/anon/Documents/ComfyUI/models/diffusion_models/qwen-image/qwen_image_bf16.safetensors     --vae /home/anon/Documents/diffusion-pipe/models/qwen-image/vae/diffusion_pytorch_model.safetensors     --text_encoder /home/anon/Documents/ComfyUI/models/text_encoders/qwen_2.5_vl_7b.safetensors     --dataset_config /home/anon/Documents/musubi-tuner/data/ruri-rocks-cfg/ruri-rocks-dataset.toml     --flash_attn --mixed_precision bf16     --timestep_sampling shift     --weighting_scheme none --discrete_flow_shift 2.2     --optimizer_type adamw8bit --learning_rate 5e-5 --gradient_checkpointing     --max_data_loader_n_workers 2 --persistent_data_loader_workers     --network_module networks.lora_qwen_image     --network_dim 16     --max_train_epochs 16 --save_every_n_epochs 1 --seed 23571113     --save_state --logging_dir /home/anon/Documents/musubi-tuner/data/ruri-rocks-logs

Using a 4090D 48GB, averaging 45GB, over 23 hours.