AI Can Turn Video Games Into Real-Life Footage Without Losing 3D Accuracy
Based on research by Dana Cohen-Bar, Ido Sobol, Raphael Bensadoun, Shelly Sheynin, Oran Gafni
A new system solves a long-standing dilemma in computer graphics by merging perfect digital control with lifelike realism, effectively erasing the fake look that usually plagues rendered scenes. While video generation models create stunning visuals, they often fail at maintaining consistent 3D geometry, whereas traditional 3D engines offer structural perfection but lack natural lighting and texture.
RealMaster bridges this gap by using specialized AI to transform game footage into photorealistic video while strictly preserving the original scene's physics and object positions. The method trains on paired datasets where realistic anchors guide intermediate frames through geometric cues, allowing the model to handle dynamic elements like moving characters that appear mid-sequence. Testing on complex Grand Theft Auto V scenarios shows the approach significantly outperforms existing editing tools, achieving a level of visual fidelity previously thought impossible for game engines.
The technology marks a major step forward for film and gaming production, enabling creators to generate high-fidelity content directly from digital blueprints without compromising structural integrity or requiring frame-by-frame manual retouching.
RealMaster: Lifting Rendered Scenes into Photorealistic Video by Dana Cohen-Bar et al., https://arxiv.org/abs/2603.23462