fudan-zvg/diffusion-square ? reverse-engineered prompt

Build me a Python app that runs this Diffusion² pipeline for generating dynamic 3D content from a single image or a short input video.

I want one simple command that takes an input folder of frames or one image, then produces synchronized multi view, multi frame outputs like the paper shows. Please make it handle the basic preprocessing automatically, like resizing, recentering, and optional background removal when needed, so I do not have to prepare the input much.

Use the pretrained video and multi view diffusion checkpoints from Hugging Face, and make it easy to point the app at a local checkpoints folder. Also include a clean way to set the camera elevation and image frame ratio from the command line.

If there is already enough in the repo to run the main inference flow, connect it up and make the instructions clear. If anything is missing, fill it in so the project runs end to end, and look up current docs online if you need to.