Skip to content

LoRA Adapters

Load and unload LoRA weights at runtime to specialize a base Sana model.

from strands_sana import sana_load_lora, sana_generate, sana_unload_loras

sana_load_lora(
    repo_or_path="user/sana-yarnart-lora",
    scale=0.8,
    model="sana-1.6b-1024",
)

sana_generate(prompt="a duck in yarn art style",
              model="sana-1.6b-1024", steps=15)

sana_unload_loras(model="sana-1.6b-1024")

Requires PEFT

pip install 'strands-sana[lora]'
# OR: pip install peft

Stacking multiple LoRAs

sana_load_lora("user/style-lora", scale=0.6, adapter_name="style")
sana_load_lora("user/character-lora", scale=0.7, adapter_name="char")
sana_generate(prompt="...")

The wrapper tracks active adapters in pipe._loaded_loras.

Sources

  • HuggingFace Hub: user/repo format
  • Local path: /path/to/lora.safetensors
  • Train your own: see Training

Train your own LoRA

from strands_sana import sana_train_lora

sana_train_lora(
    instance_data_dir="./my-photos",
    instance_prompt="a photo of sks dog",
    max_train_steps=500,
    num_processes=4,    # 4 GPUs
    dry_run=False,      # actually launch
)

→ Full training guide: Training