Disney Research and ETH Zurich Propose RenderFlow: Single-Step Neural Rendering via Flow Matching

Researchers from Disney Research|Studios and ETH Zurich have introduced RenderFlow, an end-to-end neural rendering framework that replaces conventional physically based rendering (PBR) pipelines with a flow matching paradigm. Unlike diffusion-based approaches, RenderFlow is deterministic and completes rendering in a single step, addressing the twin limitations of iterative latency and stochastic inconsistency that have hampered prior deep learning renderers. The system uses G-buffers as input and introduces a sparse keyframe guidance module to improve physical plausibility and temporal consistency, achieving near real-time photorealistic output.

The paper, accepted to CVPR 2026, also demonstrates the framework's versatility through a lightweight adapter module that repurposes the pretrained model for inverse rendering — specifically intrinsic decomposition. By bridging the efficiency of modern generative models with the precision of traditional PBR pipelines, RenderFlow represents a potentially significant step toward practical real-time neural rendering for film and VFX production workflows.