GEN3C

Overview

GEN3C is NVIDIA’s new AI video tool that knows how to work in 3D not just 2D. It takes one or more images, predicts their depth and builds a point cloud. That scene acts as a cache so every video frame lines up without weird flicker or objects jumping around.

You tell it exactly where the camera should go and it follows. Want to start inside a car and swing up to a bird’s eye shot? It’ll do that without guessing or drifting like older models tend to do.

The team from NVIDIA’s Spatial Intelligence Lab showed GEN3C at CVPR 2025. They put the code on GitHub and Hugging Face with an open model license so you can use it in personal or paid projects. They plan to add tools that can predict depth and pose straight from input so you won’t need extra steps.

People are using it for turning still photos into moving shots and making driving scenes with both eye level and aerial views.

GEN3C Camera Controls Example
--trajectory sets camera path: left (default), right, up, down, zoom_in, zoom_out, clockwise, counterclockwise.
--camera_rotation sets rotation: center_facing (default), no_rotation, trajectory_aligned.
--movement_distance changes how far the camera moves. Default is 0.3. Bigger numbers mean more dramatic moves.

VRAM Requirements
Tested on H100 and A100. For smaller GPUs, add offload flags to cut memory use. Full offloading used about 43GB during tests.