InstantHDR: Single-forward Gaussian Splatting for High Dynamic Range 3D Reconstruction

A research team has introduced InstantHDR, a novel feed-forward network that reconstructs high dynamic range (HDR) 3D scenes from standard, multi-exposure photo collections in a single forward pass. The system addresses a major bottleneck in 3D reconstruction by eliminating the need for known camera poses, pre-initialized point clouds, and the time-consuming per-scene optimization that plagues current HDR pipelines. Instead, it leverages a geometry-guided appearance modeling technique for multi-exposure fusion and a meta-network for generalizable, scene-specific tone mapping, allowing it to work with uncalibrated input images.

To train this generalizable model, the team created HDR-Pretrain, a new dataset featuring 168 Blender-rendered scenes with diverse lighting types and multiple camera response functions, filling a critical gap in available HDR training data. In comprehensive experiments, InstantHDR delivered synthesis performance comparable to state-of-the-art optimization-based HDR methods while achieving a staggering ~700x speedup in reconstruction time in its single-forward setting. Even in a post-optimization mode for maximum quality, it maintained a ~20x speed improvement, dramatically lowering the barrier to creating photorealistic 3D assets.

The implications are significant for fields like virtual production, architectural visualization, and immersive media, where capturing real-world lighting accurately is paramount. By turning a process that could take hours into one that takes seconds, InstantHDR makes high-fidelity 3D reconstruction from photographs vastly more accessible and practical for professional workflows.