yyang181/colormnet — reverse-engineered prompt

Build me a simple local demo for ColorMNet that lets me colorize a black and white video using a color reference frame. I want to open a web page, upload or choose input frames or a short video, provide the reference image, run the pretrained model, and then preview and download the colorized result.

Please make the existing app.py and test flow work cleanly, with clear setup steps for a CUDA NVIDIA GPU machine. It should create any missing folders like the model checkpoint folder, tell me where to put or download the pretrained weights, and show friendly errors if CUDA, PyTorch, or the checkpoint is missing.

Keep it focused on inference and the Gradio demo, not training. Add small example inputs if needed, make the default paths sensible, and update the README so a lightly technical person can install it, run it, and understand what to expect. Look up current docs online if you need to.