What is opsd-v?

OPSD-V is a research project from Meituan and HKUST that improves AI systems which generate long videos step by step, a technique called autoregressive video generation. These systems create a video piece by piece, and each new piece is built using memory of what came before, stored in something called a cache. Over a long video, small mistakes or weak motion in that memory can build up, making later parts of the video look worse or move less naturally. OPSD-V is a training method, not a new video generator from scratch. It takes an existing few step video generator and continues training it so it produces better long videos, without changing how fast it runs or how it generates video at the end. During this extra training, the model, called the student, keeps generating video exactly the way it normally would, building its own memory as it goes. Alongside it, a second copy of the model, called the teacher, looks at the same moments in the video but is given a cleaner memory built from real video footage instead of its own past output. The teacher then shows the student what a better result would have looked like at each of those moments, so the student learns to correct the drift that would otherwise build up over a long video. The project includes code for installing the system, downloading model checkpoints, running training, and running inference to generate new videos. The paper describing this method is available on arXiv, and there is also a project page and a set of pretrained model files hosted on Hugging Face. The authors also link a WeChat group where people can discuss the project, though that particular invite code has a listed expiration date. The project is released under the Apache 2.0 license, and the code is written in Python. It relies on tools for managing large model training across multiple computers, including memory saving techniques and support for lightweight fine-tuning methods, since training on long videos this way needs a fair amount of computing power.