OpenNovelty.org

HDR-4DGS: High Dynamic Range 4D Gaussian Splatting from Alternating-exposure Monocular Videos

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.0Download Report PDF
High Dynamic RangeGaussian SplattingMonocular 4D Reconstrution
Abstract:

We introduce HDR-4DGS, the first system for reconstructing renderable 4D high dynamic range (HDR) scenes from unposed monocular low dynamic range (LDR) videos captured with alternating exposures. To tackle such a challenging problem, we present a unified framework with two-stage optimization approach based on Gaussian Splatting. The first stage learns a video HDR Gaussian representation in orthographic camera coordinate space, eliminating the need for camera poses and enabling robust initial HDR video reconstruction. The second stage transforms video Gaussians into world space and jointly refines the world Gaussians with camera poses. Furthermore, we propose a temporal luminance regularization strategy to enhance the temporal consistency of the HDR appearance. Since our task has not been studied before, we construct a new evaluation benchmark using publicly available datasets for HDR video reconstruction. Extensive experiments demonstrate that HDR-4DGS significantly outperforms alternative solutions adapted from state-of-the-art methods in both rendering quality and speed.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces HDR-4DGS, a system for reconstructing 4D high dynamic range scenes from unposed monocular videos with alternating exposures using Gaussian Splatting. It resides in the 'Unposed Monocular Alternating-Exposure 4D HDR' leaf, which contains only two papers total (including this work). This indicates a highly sparse research direction within the broader taxonomy of 4D HDR reconstruction. The sibling paper, Mono4DGS-HDR, shares the same technical foundation, suggesting this specific problem formulation—joint pose estimation, dynamic scene modeling, and HDR synthesis from alternating exposures—is still in its early stages.

The taxonomy reveals three main branches: Gaussian Splatting-Based 4D HDR Reconstruction (where this paper sits), Learning-Based HDR Video Reconstruction (including multi-stage alignment networks and GAN-based synthesis), and Monocular 4D Dynamic Scene Reconstruction (focused on geometry without HDR). The paper's approach diverges from learning-based methods like SKFHDRNet and HDRVideo-GAN, which rely on trained networks for exposure fusion, by using explicit Gaussian representations for joint optimization. It also differs from broader monocular 4D methods like Vivid4D and DRSM, which do not address HDR synthesis or alternating-exposure input.

Among the three contributions analyzed, each was examined against a limited candidate pool: the core HDR-4DGS system (10 candidates, 1 refutable), the two-stage optimization framework (3 candidates, 1 refutable), and temporal luminance regularization (7 candidates, 1 refutable). The analysis is based on 20 total candidates from semantic search and citation expansion. The statistics suggest that while each contribution has at least one overlapping prior work among the examined candidates, the majority of candidates (9 out of 10 for the core system, 2 out of 3 for the framework, 6 out of 7 for regularization) do not clearly refute the novelty.

Given the sparse taxonomy leaf (only 2 papers) and the limited search scope (20 candidates), the work appears to occupy a relatively unexplored niche. The presence of one sibling paper and scattered refutable candidates suggests incremental refinement over closely related methods rather than a completely new problem formulation. However, the analysis does not cover exhaustive literature review, and the true novelty may depend on technical details not captured in the abstract-level comparison.

Taxonomy

Core-task Taxonomy Papers
5
3
Claimed Contributions
17
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: Reconstructing 4D high dynamic range scenes from unposed monocular alternating-exposure videos. This emerging field sits at the intersection of dynamic scene reconstruction, HDR imaging, and camera pose estimation. The taxonomy reveals three main branches that reflect different methodological emphases. Gaussian Splatting-Based 4D HDR Reconstruction leverages explicit 3D Gaussian representations to model both geometry and radiance in dynamic scenes, enabling efficient rendering and HDR fusion from alternating exposures. Learning-Based HDR Video Reconstruction focuses on neural network architectures that directly synthesize HDR frames from low dynamic range input sequences, often relying on temporal consistency and exposure bracketing patterns; representative works include SKFHDRNet[5] and HDRVideo-GAN[6], which employ deep learning to merge multi-exposure frames. Monocular 4D Dynamic Scene Reconstruction addresses the broader challenge of recovering time-varying geometry and appearance from single-camera video, with methods like Vivid4D[3] and DRSM[4] tackling motion, deformation, and view synthesis without requiring known camera poses. Within the Gaussian splatting branch, a small handful of works specifically target the unposed monocular alternating-exposure setting, combining pose estimation with HDR radiance modeling in a unified framework. HDR-4DGS[0] exemplifies this direction, jointly optimizing camera trajectories and 4D Gaussian representations to reconstruct high dynamic range dynamic scenes from raw alternating-exposure video. Its closest neighbor, Mono4DGS-HDR[1], shares the same technical foundation but may differ in how exposure alignment or temporal regularization is handled. Compared to learning-based approaches like Pix2HDR[2], which rely on training data to hallucinate HDR content, the Gaussian splatting methods offer more explicit geometric control and can adapt to novel scenes without retraining. The main open questions revolve around robustness to rapid motion, scalability to longer sequences, and the trade-off between reconstruction fidelity and computational cost when jointly solving for pose, geometry, and radiance.

Claimed Contributions

HDR-4DGS system for 4D HDR reconstruction from unposed monocular alternating-exposure videos

The authors present the first system capable of reconstructing 4D HDR scenes from monocular LDR videos with alternating exposures and unknown camera parameters. This addresses a previously unexplored challenging task in HDR novel view synthesis.

9 retrieved papers
Two-stage optimization framework with video-to-world Gaussian transformation

The authors introduce a novel two-stage optimization approach where the first stage learns HDR Gaussians in orthographic camera coordinate space without requiring camera poses, and the second stage transforms these to world space and jointly refines world Gaussians with camera parameters. This includes a video-to-world Gaussian transformation strategy based on 2D covariance invariance.

2 retrieved papers
Temporal luminance regularization strategy for HDR temporal consistency

The authors propose a temporal luminance regularization strategy using flow-guided photometric loss to align per-pixel HDR irradiance between consecutive frames. This ensures temporally consistent HDR appearance across the reconstructed video, particularly for dynamic content.

6 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

HDR-4DGS system for 4D HDR reconstruction from unposed monocular alternating-exposure videos

The authors present the first system capable of reconstructing 4D HDR scenes from monocular LDR videos with alternating exposures and unknown camera parameters. This addresses a previously unexplored challenging task in HDR novel view synthesis.

Contribution

Two-stage optimization framework with video-to-world Gaussian transformation

The authors introduce a novel two-stage optimization approach where the first stage learns HDR Gaussians in orthographic camera coordinate space without requiring camera poses, and the second stage transforms these to world space and jointly refines world Gaussians with camera parameters. This includes a video-to-world Gaussian transformation strategy based on 2D covariance invariance.

Contribution

Temporal luminance regularization strategy for HDR temporal consistency

The authors propose a temporal luminance regularization strategy using flow-guided photometric loss to align per-pixel HDR irradiance between consecutive frames. This ensures temporally consistent HDR appearance across the reconstructed video, particularly for dynamic content.

Table of Contents