Use Harmonizer for Post-Processing

Use Harmonizer for all NuRec post-processing. We recommend that you use Harmonizer instead of previous models (Difix and Fixer) for simulation and rendering pipelines.

See also

Harmonizer is open source. View the source, file issues, and contribute at github.com/NVIDIA/harmonizer.

Harmonizer

Harmonizer is a diffusion-based enhancement model designed to bridge neural reconstruction and photorealistic simulation. Its backbone is CosmosPredict2 0.6B, fine-tuned using a custom data curation pipeline for autonomous driving and large-scale robotics simulation.

Harmonizer operates downstream of reconstruction as a simulation-frame enhancer. In temporal run mode, Harmonizer conditions on previous frames during both training and inference, applying a Temporal Total Variation loss to prevent flickering and temporal drift.

Harmonizer is enabled in the gRPC server with --enable-harmonizer. See Use the NuRec gRPC API Server for configuration details.

Harmonizer provides:

• Temporally consistent video output — no flickering or frame-to-frame drift

• Cross-domain appearance matching for dynamic objects (vehicles, pedestrians) inserted from a different capture domain

• Lighting realism — shadows, color harmonization

• Production-grade simulation fidelity for real-time autonomous driving and robotics simulators