BLURD: A New Standard in Synthetic Photoreal Representation Learning Datasets
Learn about the BLURD datasets and challenges.
Fully Synthetic Dataset Creation
Merging deterministic synthetic dataset creation with data-driven generative approaches
Vary nine separate factors of variation, fully extendable to many more
Generate unlimited photorealistic images corresponing to a choosen fixed set of factors
More than just images
Every render is accompanied by a depth map, normal map, pixel colormap and set of segmentation masks for additional down stream tasks.
BLURD 3D: Unbiased Pathtraced Renders
BLURD 3D is created with high quality assets and rendered with the physically-based path traced render engine Cycles
Unbiased rendering
Unlike real-time render engines, Blender's Cycles engine does not introduce any bias into the final render. Instead Cycles accurately replicates real-world lighting and shading
High Quality 3D assets
BLURD 3D makes use of high quality 3D assets from the Blender Market place and Polyhaven
BLURD SD
A new synthetic photorealistic representation dataset
Photo-realism using Diffusion models
BLURD SD's innovative approach of combining 3D render with diffusion resulting in unparalleled photo-realism in synthetic imagery.
BLURD MASK
A new domain adaption challenge
Measure the 3D render to photorealism gap
BLURD Mask allows for the testing of diffusion based data augmentation tasks
Benchmark against real world data
Compare against a real-world dataset of human faces