Rendersynth：Blender 渲染与灯光插件

Rendersynth

An open-source Blender extension for generating synthetic datasets through procedural randomization and batch rendering. It lets you create diverse image datasets with controlled variations in object placement, material properties, lighting conditions and scene parameters, all without writing Python code.

The extension is intended for low to medium sized computer vision projects, which benefit more from ease of generation rather than the efficiency and generality of tools like BlenderProc.

Overview

The Generator provides a building block system for defining randomization pipelines directly inside Blender. You specify how objects, materials and lighting should vary across renders, then the extension outputs images with corresponding metadata suitable for training computer vision models.

You can also define named classes and assign them to Blender objects, and group multiple objects together into multi-object entities that get treated as a single labeled object.

Features

Randomization is distribution-based: each parameter can be driven by a preset distribution (normal, uniform, exponential, etc.) or by a custom distribution built visually with the node editor. Pipelines are built by concatenating operations that target object transforms, materials and lighting, and they can be saved and loaded as JSON for version control and reproducibility.

Parameters you can randomize include object position, rotation and scale, material textures and properties like roughness and metallic, visibility and occlusion, lighting intensity, temperature and color, and camera position, including locking the camera around a central point.

Rendered datasets can be exported in formats compatible with common computer vision frameworks, currently:

• Ultralytics YOLO
• COCO
• COCO Keypoints
• COCO Segmentation
• Depth maps
• Normal maps
• Point clouds

A preview living inside Blender lets you inspect the output of a pipeline run, including bounding boxes, labels and visibility statistics, directly inside Blender before committing to a full batch render. A logging system is also in progress to give better visibility into single stochastic runs and the output of constraint solvers.

查看官方完整介绍 ↗