Check your environment first
Follow the current upstream quick start for supported Python, PyTorch, CUDA, and model dependencies. Check storage capacity and available GPU memory before downloading weights. Hardware requirements depend on precision, resolution, offloading, and the chosen runtime.
Keep the first run simple
Create a separate environment, download the official checkpoint, and run the documented minimal example. Record your dependency versions. Once generation works, save the environment and then experiment with a different interface or custom workflow.
Account for the full operating cost
Compare machine rental or ownership, idle time, storage, and maintenance with hosted generation. A local system can offer flexibility, but it also needs updates, monitoring, and recovery procedures. Measure your real throughput before moving a public service.
Go to the source
Official references, checked October 21, 2026.
Official repository & quick start