Headless 3D and VFX engine for agents. Research preview.
Rendering benchmarks
Draft and final measurements are recorded in RESULTS.md. Preview and path timings there are explicitly carried forward.
The ten small scenes cover blocking, a moving car proxy, crowd, interior,
night lights, camera push-in, rack focus, street, turntable and walking animation.
Procedural GLBs in
assets/ are CC0; the animated Fox retains its attribution in
assets/README.md. The turntable downloads hash-pinned ToyCar
assets with python3 benchmarks/fetch.py. All other inputs are already included.
The car is a proxy, not a photographed car or a production vehicle asset.
Rerun
Use an installed build and a real supported GPU, then run from the checkout root:
python3 benchmarks/fetch.py
python3 benchmarks/run.py --vork vork --out /tmp/vork-benchmark
The runner resolves named profiles but forces denoise=false. By default it
selects only B scenes and previs/final/path, omits draft, and does not set seed 11.
It does not reproduce the historical denoised final8 tables or their C/D scenes.
Those measurements use build 81ae2b075 and the settings listed with each table.
It records setup, a separate per-scene warmup, five warm process
times, commands, exit codes, diagnostics and PNGs. Setup uses the existing cache;
this is not a cold-cache benchmark. Use a new VORK_CACHE_DIR to measure first use
and record that choice. --resolution 320x180 --repetitions 1 is a smoke run only.
std/previs runs are previews, with different lighting/shadows; report them
separately from final/path. No timings are projections. Inspect every image and
report failures, hardware, driver, build, cache state and quality differences.
Known failure: the helper fails during camera preparation on Blender 5.2.1: its Track To axis enum is invalid, and Python exceptions do not stop preparation reliably. The historical comparisons cannot currently be reproduced with it. Repair and validation across scenes are pending. The commands below describe the intended workflow.
To create and time Blender inputs, use the separately licensed Blender benchmark helper. It imports the same scene assets and copies the camera and lights from these scene files. Its import/save preparation is outside timing; each render launches a fresh process. Run it with system Python (the helper launches Blender):
python3 integrations/blender/benchmark.py --blender blender --out /tmp/blender-benchmark
Use --cycles-device METAL on Apple Silicon or CUDA if OptiX is unavailable.
The helper renders Blender frame 0 to match vork time 0 by default. It also
translates the supplied camera/material tracks and clip selection for full-video
runs. glTF material translation, shutter integration and renderer sampling still
differ. Re-measure both tools together; do not mix results from different builds or settings.
Full preview clips
Both runners accept --scenes B-blocking,B-push-in to select inputs and --video
to render all 240 frames. The default is every B-*.json. Run previews separately:
python3 benchmarks/run.py --profiles previs --video --out /tmp/vork-preview
python3 integrations/blender/benchmark.py --engines BLENDER_EEVEE,BLENDER_WORKBENCH --video --out /tmp/blender-preview
EEVEE uses 64 samples and ray tracing; Workbench is a studio preview without the same lighting or shadows. Encoding paths differ. Keep these qualifications next to any ratio; numerical settings do not establish equal quality. Inspect the outputs before publishing. Commands include process startup and video encoding; preparation, setup and the warmup command are recorded separately.
Million birds
The bird benchmark ships all four shot templates and the
corresponding compute programs: original birds2, then skim, dive and reveal. It
runs complete vork movies and hands actual baked flock checkpoints to Blender.
Its Blender adapters remain under integrations/blender/. Blender sparse-frame
projections and excluded simulation cost are explicit. Current draft and final
numbers are in RESULTS.md.