StreamDiffusion can process offline video frame-by-frame with img2img to produce a stylised output video

The examples/vid2vid/main.py and demo/vid2vid/app.py demonstrate batch offline video stylisation: read each frame from a source video, pass it through the StreamDiffusion img2img pipeline, write the output to a destination file. Unlike the real-time screen-capture path, vid2vid doesn’t require real-time throughput—it runs as fast as the GPU allows and stores outputs. The same StreamDiffusionWrapper is reused with mode='img2img'; the only difference is the source (file vs. live capture) and sink (file vs. display queue). This mode is useful for pre-rendering assets or testing prompt and LoRA combinations offline before a live performance.

Examples

python examples/vid2vid/main.py --input path/to/input.mp4 --output path/to/output.mp4

Assessment

A live coder wants to generate a background loop for their set by running a recorded screen session through StreamDiffusion vid2vid. Which pipeline parameters should they tune first to preserve timing from the original video?