kaist-ami/Axial-mm ? reverse-engineered prompt

Build me a PyTorch project for learning based axial video motion magnification, like the ECCV 2024 paper. I want to be able to train the model from a folder of frame data, then run inference on a video that has already been split into frames and get back an amplified output video. The app should support magnifying motion along the x axis, y axis, or an arbitrary angle, and it should let me choose between a static mode and a dynamic mode.

Please also add a mode that uses temporal filtering, with options like differenceOfIIR, butter, and fir, since that’s recommended for real videos. I’d like the script to accept a pretrained checkpoint, frame rate, frequency band, and amplification strengths, and save the results in a simple way. If anything about the original paper or common usage is unclear, look up current docs online if you need to.