HDR-NSFF: High Dynamic Range Neural Scene Flow Fields
Shin Dong-Yeon, Kim Jun-Seong, Kwon Byung-Ki, Tae-Hyun Oh
TL;DR
HDR-NSFF introduces a novel 4D spatio-temporal neural scene flow framework for dynamic HDR scene reconstruction from monocular videos, overcoming limitations of traditional 2D merging methods and achieving state-of-the-art results.
Contribution
It presents a unified end-to-end pipeline for dynamic HDR scene reconstruction using 4D modeling, incorporating exposure-invariant motion estimation and generative priors.
Findings
Recoveres fine radiance details in challenging HDR scenes
Achieves coherent scene dynamics under exposure variations
Sets new state-of-the-art in space-time HDR view synthesis
Abstract
Radiance of real-world scenes typically spans a much wider dynamic range than what standard cameras can capture. While conventional HDR methods merge alternating-exposure frames, these approaches are inherently constrained to 2D pixel-level alignment, often leading to ghosting artifacts and temporal inconsistency in dynamic scenes. To address these limitations, we present HDR-NSFF, a paradigm shift from 2D-based merging to 4D spatio-temporal modeling. Our framework reconstructs dynamic HDR radiance fields from alternating-exposure monocular videos by representing the scene as a continuous function of space and time, and is compatible with both neural radiance field and 4D Gaussian Splatting (4DGS) based dynamic representations. This unified end-to-end pipeline explicitly models HDR radiance, 3D scene flow, geometry, and tone-mapping, ensuring physical plausibility and global coherence.…
Peer Reviews
Decision·ICLR 2026 Poster
- The paper focuses on novel view synthesis for dynamic scenes with randomly exposed inputs, which is a relatively unexplored area in the current novel view synthesis field and holds considerable research value. - The pipeline of the proposed method is concise and effective, and the introduced Robust Learning Strategies effectively mitigate the sensitivity of the optical flow prior to variations in input exposure. - Experimental results demonstrate that the proposed method can handle multi-expos
1. The proposed method is built upon the NSFF framework, which relies on COLMAP to obtain camera parameters from the input images. However, when the input video exhibits large exposure variations, the quality of the camera parameters estimated by COLMAP may degrade or even cause COLMAP to fail entirely. The paper does not provide a clear solution to this issue, which may limit the practical applicability of the proposed method. 2. The paper points out that modeling HDR in 3D scenes can effective
1. It proposes a learnable tone-mapping function for HDR reconstruction. 2. It introduces a new HDR dynamic scene dataset. 3. Experimental results are better than other SOTA methods.
1. I'm wondering why this paper is conducted on NSFF methods, as there are many advanced methods in NeRFs and Gaussians in recent years. Using an outdated baseline somehow weakens the contributions of the paper. 2. The writing of the method section includes so many details on pretrained model selections, whereas the main contribution about the dataset and the learnable tone mapping is relatively introduced far behind. This may mislead the reader's judgment of an important part of the paper.
The paper is well-structured, and the experiments are fairly comprehensive. Reconstructing dynamic HDR radiance fields is also an interesting and meaningful idea.
1. The visual quality of the reconstructed images is not satisfactory; noticeable blurriness can be observed in several results (e.g., Figure 1 and Figure 6). Could the authors provide an explanation for this issue? 2. Long exposure tends to introduce motion blur (e.g., in the Big Jump scene), while short exposure often leads to noise. How do the authors address this trade-off in their method? 3. When comparing with other HDR reconstruction methods, was the same tone-mapping function used for
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsImage Enhancement Techniques · Generative Adversarial Networks and Image Synthesis · Advanced Vision and Imaging