Comfy-Org/ComfyUI ? reverse-engineered prompt

I want to build a powerful app for creating AI art, videos, and audio. The main interface shouldn't be a simple form with text boxes; it should be a visual, node-based system. I'm thinking of a canvas where you can drag and drop different blocks—like a model loader, a text prompt input, a sampler, an upscaler—and then connect them with wires to build a custom generation pipeline, like a flowchart.

The system needs to be really flexible, supporting all the modern diffusion models like SDXL and SD3, but also stuff for video like Stable Video Diffusion and audio. It should handle ControlNets, LoRAs, and inpainting. It's important that it's smart about performance, only re-running the parts of the flowchart that have changed. A killer feature would be saving the entire workflow directly into the generated image file, so I can just drag and drop a PNG to load the exact process I used to create it. It should be built with Python and PyTorch.