What is rino?

RINO is a research project from a team at Johns Hopkins University, UC Santa Cruz, Carnegie Mellon, and Rice University, released alongside a paper called Let RGB Be the Language of Vision. It is the code used to reproduce and test the ideas in that paper, and it focuses on a specific idea for computer vision, the field of getting computers to understand and generate images. Normally, different vision tasks such as estimating depth in a photo, detecting objects, or figuring out a person's pose each need their own specialized model built for that exact job. RINO's core idea is to treat every one of these tasks as if it were simply editing an image, expressing both the input and the output as ordinary RGB color images. Instead of building a new model per task, it takes an existing, unmodified image editing AI model, already trained by someone else, and uses it as-is to handle understanding tasks by rendering their output, like a depth map or a segmentation mask, as a picture, and also to handle generation tasks that take a picture as their starting condition. The project supports three existing open-source image editing models as interchangeable backends: Qwen-Image-Edit, FireRed-Image-Edit, and LongCat-Image-Edit. None of these are modified or retrained, they are used purely as black boxes. Each vision task the paper tests lives in its own folder with its own evaluation code and uses officially copied scoring code so results can be fairly compared to prior published work. To try it out, a user installs the Python requirements and can run two included demo scripts that chain together, one that turns a photo into a grayscale depth map and a second that takes that depth map plus a text caption and generates a new photo from it. Model weights download automatically from Hugging Face the first time the scripts run.