VisionBlender is an open-source Blender add-on hosted on GitHub, with about 530 stars under the Apache-2.0 license and updates in 2025. It turns Blender into a synthetic dataset generator for computer vision, producing rendered sequences together with perfect ground truth annotations.
Collecting and hand-labeling real imagery is slow and expensive, so VisionBlender lets you build 3D scenes in Blender and export monocular or stereo video sequences alongside the ground truth maps that vision models and benchmarks need. Supported outputs include depth, disparity, segmentation masks, surface normals, optical flow, object pose, and camera parameters. The add-on adds a VisionBlender panel to Blender’s Output Properties, works with the fast Eevee renderer by default, and uses the Cycles engine when segmentation masks or optical flow are required.
Because every frame is rendered from a 3D scene, the ground truth is exact by construction rather than estimated by human annotators. That makes the add-on useful for benchmarking depth, disparity, and optical flow algorithms, where accurate real-world ground truth is nearly impossible to capture, and for augmenting training sets with controlled variations of viewpoint, lighting, and object placement. Each object simply needs a pass index other than zero for segmentation output.
VisionBlender is free and open source under the Apache-2.0 license. It runs inside Blender, which is also free, so the only cost is your own rendering hardware and time.
It is aimed at computer vision researchers and ML engineers who need labeled training or benchmark data with exact ground truth, especially for depth estimation, optical flow, segmentation, and pose estimation experiments.