Video color grading is a critical post-production process, transforming raw, log-encoded footage into emotionally resonant cinematic visuals. Existing automated methods often function as static, black-box executors, directly outputting edited pixels without offering the interpretability or iterative control crucial for professionals.
We introduce LumiVideo, an intelligent agentic system designed to mimic the cognitive workflow of professional colorists. It operates through four distinct stages: Perception, Reasoning, Execution, and Reflection. Given only raw log video, LumiVideo autonomously generates a cinematic base grade by analyzing the scene's physical lighting and semantic content.
The core Reasoning engine leverages an LLM's internalized cinematic knowledge, synergized with a Retrieval-Augmented Generation (RAG) framework, and employs a Tree of Thoughts (ToT) search. This combination effectively navigates the complex, non-linear color parameter space. Crucially, LumiVideo does not generate pixels; instead, it compiles the deduced parameters into industry-standard ASC-CDL configurations and a globally consistent 3D LUT, analytically guaranteeing temporal consistency.
An optional Reflection loop further enhances the system, allowing creators to refine the color grade through natural language feedback. To facilitate evaluation, we also introduce LumiGrade, the first log-encoded video benchmark specifically designed for automated grading systems. Experimental results demonstrate that LumiVideo achieves quality comparable to human experts in its fully automatic mode, while also providing precise iterative control when directed, significantly advancing video post-production capabilities.