Skip to content

Rendersynth

Current Version 1.0.0 , Tested on Blender 4.5.0

Rendersynth is a Blender extension for generating synthetic image datasets with automatic annotations for computer vision tasks. It integrates directly into Blender's 3D viewport sidebar and drives the full generation pipeline: randomization, rendering, labeling, and export, without leaving the application or requiring any external tooling.

The extension lets you define a pipeline of randomization stages (object transforms, material swaps, lighting adjustments, camera movement, and more), configure a labeling scheme, and produce a batch of rendered images with corresponding annotations in standard formats such as Ultralytics YOLO, COCO, Pascal VOC, and CVAT XML.

What it does

At its core, Rendersynth separates three concerns cleanly: what to randomize, how to label the result, and where to write the output.

Randomization is expressed as an ordered sequence of "pipes," each targeting a specific Blender property: object position, rotation, scale, visibility, material, shader node value, texture, light power, focal length, and more. Each pipe draws samples from a configurable probability distribution (uniform, Gaussian, Beta, Bernoulli, categorical, and several multivariate variants). For advanced use, any distribution parameter can be driven by a node graph in Blender's own node editor rather than a fixed preset.

Labeling is built on a camera-space ray casting pass that identifies which objects are visible in a given render and to what degree they are occluded. Objects are mapped to user-defined classes either directly, via naming rules, or via collection membership. Multi-object entities can be defined so that several scene objects are jointly annotated as a single label. Geometry is currently extracted as axis-aligned bounding boxes, with a convex hull extractor also present.

Output is written through a strategy pattern: each supported annotation format implements a common interface, and the writer calls into it per-image or per-batch depending on the format's requirements. Images are saved via Blender's own render pipeline; no external image library is involved.

A preview mode renders a single sample from the pipeline and overlays the annotation geometry, class labels, and timing statistics directly on the rendered image all inside Blender, with no external viewer.

Capabilities at a glance

  • Pipeline-based randomization of object transforms, materials, textures, shader properties, lights, and cameras
  • Arbitrary probability distributions configurable from the sidebar, with an optional visual node editor for complex distributions
  • Class assignment via direct mapping, naming rules, or Blender collections
  • Multi-object entity support for jointly-annotated composite objects
  • Visibility estimation via 3D bounding box projection vs. ray-traced visible area
  • Annotation export in Ultralytics YOLO, COCO (detection), Pascal VOC, and CVAT XML formats
  • In-Blender annotated preview with timing diagnostics
  • Pipeline serialization to and from JSON for reuse and version control
  • Optional structured logging of stochastic outputs for reproducibility analysis
  • Zero external Python dependencies: everything runs on Blender's bundled Python environment

Changelog

[1.0.0] - Current release - First stable public release.

[0.9.1] - 13 May 2026 - Implementation of randomization stages for material, lights, and cameras. - Implementation of color randomization for shaders.

[0.9.0] - 29 April 2026 - Implementation of COCO polygon labels. - Implementation of CVAT XML format. - Implementation of Pascal VOC XML format. - Fixed a bug where the YOLO format would not generate the aggregated .yaml and data.txt.

Contributing

Rendersynth is open source under the MIT license. Contributions or bug reports, feature proposals, new labeling format implementations, or additional randomization stages are all welcome via the GitHub repository.

If you wish to add a new stage to the pipeline, before opening a pull request, please check that any new operator is registered in operators/names.py using the Labels enum (rather than as a raw string), and that any new pipe type has a corresponding entry in OperationRegistry, OperationDrawerRegistry, and PipeSchemaRegistry. The architecture documentation describes these registries in more detail.

For questions or discussion, please open an issue or contact my email.