H-oliday/SwiftVR ? reverse-engineered prompt

Build me this SwiftVR project so I can restore low quality videos into clean high quality ones from the command line and from Python.

I want to be able to point it at an MP4 or an image folder, choose an upscale factor, and get back a restored video or PNG sequence. Please make the package installable, include the inference script, and wire up a simple pipeline object that can load pretrained weights from a checkpoint folder. It should also support streaming chunk by chunk so I can process long videos without waiting for the whole file to finish.

Use the README as the source of truth for how it should work, and make sure the example commands and Python usage actually run. If you need to check the latest PyTorch or diffusion docs online to get the setup right, go ahead. Keep the code clean and practical, with sensible defaults for things like clip length, output quality, and bfloat16 on GPU.