Meta Researcher 남길주: Image or Video Output May Not Require 3D Modeling; Robotics Still Needs Spatial Understanding
According to ZDNet Korea, Meta Reality Labs researcher 남길주 argued at AI Tech Day 2026 at Korea University that producing an image or video does not necessarily require building a complex 3D model. Drawing on roughly five years of his own 3D hair-reconstruction research, he discussed possible directions for 3D graphics expertise in robotics and autonomous driving, where spatial and physical reasoning remain necessary.
According to ZDNet Korea, Meta Reality Labs researcher 남길주 sees generative AI as changing the role of 3D modeling in video production. In a talk at AI Tech Day 2026 held at Korea University, he offered a conditional judgment: when the final output is an image or video, building a 3D model is not necessarily required, because generative AI can produce visuals without individually performing the traditional modeling steps. He used this to discuss new directions for 3D graphics in robotics, autonomous driving and other physical-world tasks; the report does not establish this across every film and game production workflow.
Traditional computer graphics works by first constructing the 3D shape of characters and objects, then computing movement and light reflections to render the scene. Generative AI skips these steps — it learns from vast amounts of image and video data and produces new visuals directly from text descriptions. Early AI video mainly generated faces; later it learned to handle complex backgrounds and light reflections. OpenAI's Sora can already realistically reproduce human gait and surrounding environments, and recently techniques have emerged that control a character's facial expressions and body movements.
The report also recounts 남길주's own research experience. He has been researching computer vision and 3D graphics at Meta Reality Labs since 2020, and spent roughly five years working on 3D hair reconstruction. He eventually changed his approach: instead of hand-designing complex graphics pipelines, he applied only basic knowledge and let AI learn the rest from data. He described the result as better than the original method.