wtf is defringe?

Defringe is a tool for removing green and purple color fringing from videos, the faint colored halos that sometimes appear around high contrast edges in footage. It works only on the color channels, leaving the underlying brightness and detail of the image untouched. The approach treats fringe like a shadow cast by a light source. The tool first finds the casters, the bright or saturated spots responsible for the fringe, then locates the nearby colored halo those spots produce, and finally pulls that stray color back toward the clean, natural tone around it. Green fringe tends to come from warm, saturated sources like red or purple areas, while purple fringe comes from overly bright, blown out highlights. Because purple fringe can itself cause green fringe nearby, the tool is designed to remove green fringe first, then purple, so nothing gets left behind without its source. There are two versions of the core algorithm that are kept in sync with each other and tested against one another: one written in plain numpy, and a second built with PyTorch so it can be exported as an ONNX model and run efficiently elsewhere, including on a GPU through a provided Colab notebook. The included app is a five tab visual tool built with Gradio. You load a video clip, then tune the green and purple correction passes using color wheel controls while overlays show exactly what the algorithm is detecting and correcting. A temporal tab lets you scrub through corrected frames and check for flicker over time. The final tab bakes your tuned settings into a portable ONNX model you can then run across an entire video. To try it, you need the uv Python tool and ffmpeg installed. Once those are in place, a single command downloads a compatible Python version and all the dependencies, then launches the tuning app in your browser.