Feature Preview: Neural Rendering
Neural rendering adds photoreal detail and new environmental diversity on top of Bifrost's 3D engine, without breaking your labels.
Simulation gives you control: exact scenarios, exact sensors, pixel-perfect labels. Neural rendering adds a generative layer on top - pushing realism deeper and diversity wider than what you'd author by hand, while keeping every output anchored to the 3D scene underneath.
Neural Rendering is now available in preview on our credits plan.
What's New
- Go deeper on realism. Add granular detail like weathering, surface wear, and material variation to boost photorealism across your target vessels.
- Go wider on diversity. Generate new environments, lighting, and weather beyond your existing dataset.
- 3D scene grounding. Outputs are anchored to Bifrost's 3D engine, significantly reducing hallucinations and geometry drift.
- Annotation consistency. Generated outputs stay aligned with your existing labels and metadata - no manual relabeling.
- Scalable generation workflow. Iterate in the playground, then queue batch rendering jobs to expand across locations, conditions, and sensor configurations.
Where it hits hardest
The highest-impact uses we're seeing are environmental: varying lighting, weather, backgrounds, and ocean appearance on top of simulated scenes, plus subtle camera perturbations that stress a model the way real deployments do. One simulated scenario becomes dozens of visually distinct training samples - same geometry, same labels, new conditions.
Grounded, not hallucinated
Generative models on their own drift: geometry warps, objects appear that were never in the scene, and labels stop matching pixels. Bifrost's neural rendering is grounded in the 3D engine, so the scene's structure - vessel positions, horizon, occlusions - stays fixed while appearance varies. Your annotations remain valid on every generated frame.
From playground to batch
Start in the playground: pick a rendered frame, dial in the variation you want, and iterate until the output looks right. Then queue batch rendering jobs to apply it at dataset scale across locations, conditions, and sensor configurations.
Neural rendering is improving continuously - we actively adopt state-of-the-art models as they become available, so output quality keeps climbing without any change to your workflow.
See it on your data
Want to preview neural rendering on your own target vessels and conditions?