Original title: Can LLMs Design Video Coding Tools? A Case Study on Planar Mode
Yingwen Zhang, Meng Wang, Liqiang He, Shiqi Wang · September 2026
Video compression makes video files smaller, but designing the tools inside it is hard because the algorithms are tightly connected. Researchers tested whether large language models can design these tools. They focused on Planar modeA technique in video coding that predicts pixel values based on a plane to compress data more efficiently., an established tool that predicts pixel values in video frames. To test this, the researchers set up a loop: the model generates new prediction code, video encoder trials test how well it compresses video, and the model refines its designs using that feedback. The results show this works. When tested in a fast video encoder, the model's new mode beat the default version, saving 0.18% in bitrate with 0.4% complexity overhead. It also showed compression gains in a more advanced video model under a low-resolution setting, proving language models can design real video tools.
Imagine you're trying to perfect a cake recipe but you don't know the exact amounts of sugar and flour. Each time you bake, the cake might be too sweet or too dry, so you change a little and try again based on the taste. In this paper, the AI is like a baker experimenting with ingredients, but instead of cake, it's designing parts of video compression tools. It generates code, tests how well it compresses videos, and improves based on the results.
Video coding tools affect how every online video is compressed and streamed, impacting data usage and quality. If AI can create better tools, it could lead to smaller file sizes and higher-quality videos without needing more bandwidth, benefiting viewers and streaming services.
The study only tests on one video coding tool, Planar mode, and within specific settings like faster presets and low-resolution. It does not establish whether the approach generalizes to other coding tools or higher resolutions. The gains are preliminary and may not hold in all real-world scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Can LLMs Design Video Coding Tools? A Case Study on Planar Mode — Yingwen Zhang, Meng Wang, Liqiang He et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.