HDR-4DGS: Dynamic HDR Novel View Synthesis

• HDR-4DGS is a 4D Gaussian Splatting architecture for HDR dynamic novel view synthesis that models time-varying scene radiance with a tailored dynamic tone mapper.
• The method fuses a native 4D Gaussian scene representation with a Dynamic Radiance Context Learner to adaptively map latent HDR radiance to observed LDR images under varying exposures.
• Empirical results indicate that HDR-4DGS enhances temporal coherence and inference speed while outperforming static tone mapping schemes in capturing dynamic HDR details.

Searching arXiv for the named method and closely related HDR dynamic Gaussian papers. Searching arXiv for HDR-4DGS. HDR-4DGS is a 4D Gaussian Splatting architecture for HDR Dynamic Novel View Synthesis (HDR DNVS), a task defined as learning a time-varying HDR 3D model from multi-view, time-varying LDR observations so that HDR images can be rendered at arbitrary viewpoints and arbitrary times. Its central claim is that dynamic scenes require not only a space-time scene representation, but also a temporally adaptive mapping between HDR radiance and observed LDR intensities. Accordingly, HDR-4DGS combines a native 4DGS backbone with a Dynamic Tone Mapper (DTM) that conditions HDR-to-LDR translation on exposure time and temporal radiance context, with the stated goal of maintaining temporal radiance coherence while preserving spatially accurate color translation (Zhang et al., 26 Sep 2025).

1. Task formulation and conceptual basis

HDR-4DGS is introduced for HDR DNVS, formalized as learning a function

$$\mathcal{F} : (t', V') \rightarrow \mathbf{I}^{h}_{t',V'}$$

that outputs an HDR image $\mathbf{I}^{h}_{t',V'}$ for any queried timestamp $t'$ and viewpoint $V'$. Training data are multi-exposure LDR images

$$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$

where each frame $\mathbf{I}_t^l$ has an exposure time $e_t$ chosen from

$E = \{e_1,\dots,e_P\},$

and a camera viewpoint $V_t$ selected from

$\mathbf{V} = \{V_1,\dots,V_Q\}.$

The formulation also permits coupled HDR supervision,

$\mathbf{I}^{h}_{t',V'}$0

The method is positioned against three classes of insufficiency. First, static HDR novel-view methods implicitly assume that all scene elements remain stationary and therefore do not model temporal radiance variation. Second, dynamic LDR methods such as conventional 4DGS model spatiotemporal scene structure but do not recover HDR radiance or exposure-dependent image formation. Third, a single static tone-mapping function is treated as inadequate for dynamic scenes, because illumination, object motion, and scene radiance distributions evolve over time. The paper therefore separates the problem into two coupled parts: a 4D Gaussian scene representation that models geometry, visibility, and time-varying appearance, and a dynamic tone-mapping mechanism that maps latent HDR radiance to observed LDR intensities under changing exposure and temporal context (Zhang et al., 26 Sep 2025).

A common simplification is to describe HDR-4DGS as ordinary 4DGS with HDR-valued colors. The paper argues against that reading. Its stated novelty is not merely extending the color range, but explicitly connecting HDR and LDR domains through a dynamic tone mapper that adapts to temporal radiance statistics. This suggests that the method should be understood as a factorization of dynamic scene modeling and dynamic photometric adaptation, rather than as a purely radiance-range extension.

2. Native 4D Gaussian scene representation

HDR-4DGS adopts the 4DGS rendering formalism

$\mathbf{I}^{h}_{t',V'}$1

where $\mathbf{I}^{h}_{t',V'}$2 are pixel coordinates, $\mathbf{I}^{h}_{t',V'}$3 is time, $\mathbf{I}^{h}_{t',V'}$4 is the temporal marginal probability that Gaussian $\mathbf{I}^{h}_{t',V'}$5 contributes at time $\mathbf{I}^{h}_{t',V'}$6, $\mathbf{I}^{h}_{t',V'}$7 is its conditional spatial probability, $\mathbf{I}^{h}_{t',V'}$8 is opacity, and $\mathbf{I}^{h}_{t',V'}$9 is color. This is a native space-time formulation rather than a canonical-deformation field written in separate stages.

Each Gaussian is parameterized by a 4D mean

$t'$0

and a covariance

$t'$1

where $t'$2 is an extended 4D rotation matrix and $t'$3 is a scaling matrix. The temporal marginal is explicitly modeled as

$t'$4

The paper states that 4DGS uses 4D spherical harmonics (4DSH), built by combining traditional spherical harmonics with Fourier series, to model time-varying, view-dependent color. HDR-4DGS extends this color space from LDR to HDR, so the Gaussian appearance parameters are interpreted as HDR radiance attributes rather than bounded display colors.

This representation assigns temporal variation at two levels. The first is in the scene itself: 4D Gaussian support and 4DSH encode time-varying geometry and appearance. The second is in the observation model: the same latent HDR radiance is mapped to different LDR images depending on exposure and temporal radiance context. This division is central to the method’s claim that dynamic HDR synthesis requires more than a dynamic geometry model.

3. Dynamic Tone Mapper and temporal radiance context

The Dynamic Tone Mapper maps per-time HDR Gaussian colors to LDR colors as

$t'$5

where $t'$6 denotes HDR colors at time $t'$7, $t'$8 denotes the corresponding LDR colors, and $t'$9 is exposure time. The motivation is explicitly temporal: a static HDR-to-LDR mapping is considered insufficient because the radiance distribution of a dynamic scene changes over time.

To condition tone mapping on scene brightness history, the method constructs a radiance bank. For each timestamp $V'$0, it computes

$V'$1

the mean HDR color over all Gaussian points at that time. A temporal window of previous radiance signatures,

$V'$2

is then processed by the Dynamic Radiance Context Learner (DRCL):

$V'$3

The paper states that DRCL can be implemented with RNN, LSTM, GRU, or Transformer variants, and uses GRU by default.

The actual adaptive mapping is

$V'$4

where $V'$5 is a per-channel learnable tone-mapping function and $V'$6 denotes concatenation. The term $V'$7 is $V'$8, i.e. exposure-scaled HDR radiance in log domain. The paper states that $V'$9 adopts the same structure as HDR-GS, but is now conditioned on dynamic temporal context.

This architecture is the principal answer to the claim that temporal radiance coherence cannot be maintained by a fixed camera-response surrogate. Rather than imposing a direct pairwise smoothness loss on LDR outputs, the method builds temporal dependence into the observation model itself. A plausible implication is that temporal coherence emerges from shared context-conditioned tone mapping, not only from standard reconstruction losses.

4. Supervision, losses, and optimization

HDR-4DGS is trained jointly over the 4D Gaussian scene parameters, HDR appearance, the DTM parameters $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$0, and the DRCL sequence model. Its total objective is

$$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$1

where $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$2 is LDR-domain supervision and $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$3 is an HDR reconstruction term. The paper states that $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$4 if HDR ground truth is unavailable, so the system supports both LDR-only training and joint LDR+HDR training.

A distinctive design choice is the use of both a tone-mapped HDR rendering and a direct LDR rendering. The tone-mapped HDR image is written

$$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$5

where $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$6 is the HDR image rasterized by the HDR Gaussian model. The paper also supervises a direct Gaussian-rendered LDR image $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$7, and defines a reconstruction loss

$$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$8

with $$\mathbf{I}^l = \{\mathbf{I}_1^l, \cdots, \mathbf{I}_T^l\},$$9. The LDR supervision is then

$\mathbf{I}_t^l$0

HDR supervision is applied in a compressed domain:

$\mathbf{I}_t^l$1

with $\mathbf{I}_t^l$2-law compression

$\mathbf{I}_t^l$3

This avoids direct loss computation on unbounded HDR values while remaining tied to the latent HDR representation.

The implementation details reported are concise: optimizer Adam, DTM learning rate $\mathbf{I}_t^l$4, temporal context dimension $\mathbf{I}_t^l$5, default DRCL GRU, and the same training iterations as 4DGS on a single NVIDIA RTX 4090. The paper also states that adding HDR supervision can reduce model redundancy and improve inference speed, which it attributes to reduced need for redundant Gaussians.

Two misconceptions are addressed by the objective itself. First, HDR ground truth is not required, because LDR-only optimization is explicitly supported by setting $\mathbf{I}_t^l$6. Second, the DTM is not treated as a post hoc display transform; it is optimized jointly with the scene representation and appears in the supervision path.

5. Empirical behavior, datasets, and ablations

The paper introduces two benchmarks. HDR-4D-Syn contains eight synthetic dynamic scenes—Airplane, Deer, Hook, Jump, Lego, Mutant, Standup, and Tank—with synchronized multi-exposure video sequences and re-synthesized HDR ground truth. HDR-4D-Real contains four real dynamic indoor scenes—Bed, Excavator, Tank, and Toys—captured with six synchronized iPhone 14 Pro devices for some scenes, with HDR ground truth generated by UltraFusion. For synthetic scenes, Gaussians are randomly initialized with $\mathbf{I}_t^l$7 primitives; for real scenes, they are initialized from a dense reconstructed point cloud and the images are downsampled by a factor of 4 (Zhang et al., 26 Sep 2025).

On HDR-4D-Syn under LDR-only supervision, HDR-4DGS reports HDR performance of 25.88 PSNR / 0.865 SSIM / 0.076 LPIPS and LDR performance of 33.16 PSNR / 0.949 SSIM / 0.055 LPIPS, with 69.38 min training and 40.80 fps inference. Under LDR+HDR supervision, HDR-4DGS$\mathbf{I}_t^l$8 reports 30.40 PSNR / 0.914 SSIM / 0.097 LPIPS on HDR and 30.69 / 0.927 / 0.097 on LDR, with 48.63 fps inference. On HDR-4D-Real, HDR-4DGS reports 14.50 / 0.884 / 0.200 on HDR and 26.88 / 0.825 / 0.221 on LDR under LDR-only supervision, while HDR-4DGS$\mathbf{I}_t^l$9 reports 25.13 / 0.909 / 0.162 on HDR and 30.69 / 0.927 / 0.097 on LDR under LDR+HDR supervision. The paper also notes that HDR-HexPlane$e_t$0 reaches 33.03 / 0.904 / 0.192 on real-scene HDR metrics, and argues that this is partly a PSNR artifact because PSNR can prefer overly smooth outputs.

The runtime comparison is part of the method’s positioning. The paper states a speed advantage over HDR-HexPlane of up to 36× faster inference on synthetic scenes and up to 200× faster on real scenes. Reported inference speeds include 40.8 fps on HDR-4D-Syn under LDR-only supervision, 48.63 fps with LDR+HDR supervision, and 35.27 fps on HDR-4D-Real under LDR-only supervision.

The ablation study centers on the DTM. Replacing the dynamic tone mapper with Reinhard, Durand, or an MLP (HDR-GS style) produces lower HDR scores on HDR-4D-Syn: 22.10 / 0.812 / 0.210, 22.85 / 0.825 / 0.195, and 23.92 / 0.841 / 0.142, respectively, versus 25.88 / 0.865 / 0.076 for the full method. Removing the additional pixel-level supervision reduces performance from 25.88 / 0.865 / 0.076 to 24.85 / 0.853 / 0.169. DRCL ablation yields 25.63 / 0.847 / 0.100 for RNN, 25.53 / 0.845 / 0.106 for LSTM, 25.88 / 0.865 / 0.076 for GRU, and 25.06 / 0.817 / 0.101 for Transformer. Varying the temporal window length gives 24.74 / 0.852 / 0.092 for $e_t$1, 24.76 / 0.851 / 0.098 for $e_t$2, 25.88 / 0.856 / 0.076 for $e_t$3, and 24.29 / 0.825 / 0.094 for $e_t$4.

These results support the paper’s central claim that the main gain does not come from replacing HexPlane with Gaussians alone, but from combining a native 4DGS scene model with a dynamic, context-aware tone-mapping module. They also suggest that HDR-4DGS is most effective when temporal radiance history is long enough to be informative but short enough to avoid redundant or noisy context.

6. Relation to adjacent HDR and dynamic Gaussian methods, and stated limitations

HDR-4DGS occupies a specific point in the broader design space. HDRGS introduces a static 3D HDR radiance field from multi-view, multi-exposure LDR images by redefining Gaussian color as radiance, adding an asymmetric grid for tone mapping, and using a coarse-to-fine optimization strategy (Wu et al., 2024). HDRSplat instead trains static 3DGS directly on 14-bit linear raw images with a linear HDR-space loss and rasterization tuning for low-light scenes (Singh et al., 2024). HDR-NSFF formulates dynamic HDR reconstruction from alternating-exposure monocular videos as a representation-agnostic 4D problem and explicitly states compatibility with both neural radiance fields and 4DGS backends, adding learnable tone mapping, DINO-based motion, and a generative prior (Dong-Yeon et al., 9 Mar 2026). Mono4DGS-HDR targets unposed monocular alternating-exposure videos with a two-stage optimization: first learning HDR video Gaussians in orthographic camera-coordinate space, then transforming them into world space and jointly refining world Gaussians with camera poses, together with temporal luminance regularization (Liu et al., 21 Oct 2025). This suggests that HDR-4DGS is neither the only HDR Gaussian dynamic model nor a direct extension of the monocular alternating-exposure line; rather, it is a multi-view dynamic HDR NVS method that emphasizes native 4DGS plus dynamic tone mapping.

The paper’s limitations are explicit. It builds on 4DGS, which was not originally designed specifically for HDR. Temporal coherence is modeled through a fixed-length context window rather than a fully adaptive temporal mechanism. The radiance bank uses global mean HDR color statistics, which the paper identifies as a compact but coarse summary of scene radiance. Dynamic boundaries can blur when foreground and background have similar appearance. Real HDR supervision generated via 2D methods may lack multi-view consistency, affecting both evaluation and training stability. The paper also notes that extreme illumination effects may still be imperfectly modeled because the base 4DGS representation is not itself HDR-specific.

These limitations clarify the method’s scope. HDR-4DGS does not attempt a physically complete camera-and-illumination model, nor does it introduce local spatially varying radiance-context adaptation. Its claim is narrower and more architectural: a native 4D Gaussian representation becomes markedly more effective for HDR dynamic novel-view synthesis when paired with a temporally adaptive tone mapper, a radiance bank, and joint HDR/LDR supervision.