Orange-3DV-Team/MoCha ? reverse-engineered prompt

Build me a simple Python app for video character replacement like the MoCha project described in the paper.

I want to give it a source video, a mask for the character in the first frame, and one or two reference images of the new character, then have it generate a replacement video that keeps the motion and scene context while swapping in the new person or character. Please make it work as an inference tool first, with a clear way to run it from the command line on example files and on my own inputs from a CSV or similar simple data file.

It should be able to handle at least one reference image, and if a second one is missing, it should still run. Please include sensible defaults, easy setup instructions, and output the generated videos into a results folder. If you need to look up current model or dependency docs online, go ahead.