jjihwan/liteframe

LiteFrame is the official GitHub repository for a research paper titled LiteFrame: Efficient Vision Encoders Unlock Frame Scaling in Video LLMs. The work comes from a team at Google DeepMind together with Seoul National University, with Jihwan Kim listed as the first author and other authors including Nikhil Parthasarathy, Danfeng Qin, Junhwa Hur, Deqing Sun, Bohyung Han, Ming-Hsuan Yang, and Boqing Gong. The README's own one-sentence summary calls the project a highly efficient video encoder for Video Large Language Models that aims to unlock scalable, long-form video understanding by addressing inefficiencies in both the language model and the Vision Transformer (ViT). In other words, the paper is about making it cheaper and more practical to feed many frames of a video into a model that combines vision and language, rather than only being able to look at a handful of frames. It is important to be plain about the current state of the repository. The README contains a clearly marked note that the code and model weights will be released soon. As of the README's news entry dated 2026.05.18, only the paper itself has been posted to arXiv. There is a 1-minute overview video linked in the README and a project page hosted on the first author's site, but no runnable training or inference code is present yet. Because the README is short and almost entirely about author credits, paper links, and the planned release, there is no install guide, no usage example, no benchmark numbers, and no description of the LiteFrame architecture itself in the text shown here. The repository acts as a placeholder that lets people cite the paper and watch for the upcoming code drop. A BibTeX citation block is included for researchers who want to reference the work in their own papers. The arXiv preprint number is 2605.17260, and the project page is at jjihwan.github.io/projects/LiteFrame. Anyone interested in actually running LiteFrame will need to wait for the authors to publish the code and weights.