LumiVideo: An Intelligent Agentic System for Video Color Grading

A research team led by Yuchen Guo has published a paper on LumiVideo, an intelligent agentic system designed to automate the complex, artistic process of video color grading. Unlike existing 'black-box' methods that directly output pixels, LumiVideo is structured as a cognitive agent that replicates a professional colorist's workflow through four distinct stages: Perception, Reasoning, Execution, and Reflection. Given raw, log-encoded footage, it first analyzes the scene's physical lighting and semantic content. Its core Reasoning engine then synergizes a large language model's (LLM) cinematic knowledge with a Retrieval-Augmented Generation (RAG) framework, using a Tree of Thoughts (ToT) search strategy to intelligently navigate the vast, non-linear space of color grading parameters.

Crucially, LumiVideo does not generate final pixels. Instead, in the Execution stage, it compiles its deduced parameters into industry-standard ASC-CDL (American Society of Cinematographers Color Decision List) configurations and a globally consistent 3D LUT (Look-Up Table). This analytical approach guarantees temporal consistency across video frames, a major challenge for automated systems. An optional Reflection loop allows creators to provide natural language feedback (e.g., 'make it warmer') for iterative refinement. The team also introduced LumiGrade, the first benchmark for evaluating automated grading on log-encoded video. Experiments indicate the system can approach human expert quality in fully automatic mode while providing the precise, interpretable control professionals demand.