---
title: HDR Dynamic Novel View Synthesis
url: https://www.emergentmind.com/topics/hdr-dynamic-novel-view-synthesis-hdr-dnvs
type: topic
---

# HDR Dynamic Novel View Synthesis

Searching arXiv for recent papers on HDR dynamic novel view synthesis and closely related HDR NVS methods.
arXiv search: HDR dynamic novel view synthesis Gaussian Splatting NeRF single exposure events
HDR Dynamic Novel View Synthesis (HDR DNVS) denotes the problem of learning a high dynamic range 3D or 4D scene representation from low dynamic range observations so that photorealistic views can be rendered from arbitrary viewpoints, and in the dynamic setting, at arbitrary time instances as well. In the formulation introduced for HDR DNVS, the additional dimension “Dynamic” emphasizes the necessity of jointly modeling temporal radiance variations alongside sophisticated 3D translation between LDR and HDR [2509.21853]. Across the literature, the topic spans exposure-calibrated neural radiance fields, dynamic radiance representations, event-assisted deblurring, low-light intrinsic decomposition, and Gaussian-splatting-based real-time systems [2111.14451] [2401.06052] [2601.15475].

## 1. Problem formulation and image formation

The foundational physical model used throughout HDR novel view synthesis is that an observed LDR pixel is a nonlinear function of scene radiance and exposure. HDR-NeRF formulates this as
$$
Z = f(H \Delta t),
$$
where $H$ is irradiance, $\Delta t$ is exposure time, $f(\cdot)$ is the camera response function, and $Z$ is the pixel value [2111.14451]. The same paper replaces the usual NeRF color field with a radiance field and a learned tone mapper, so that the rendered LDR color is
$$
c(\mathbf{r}, \Delta t) = g\left(\ln \mathbf{e}(\mathbf{r}) + \ln \Delta t\right),
$$
while the HDR image is rendered directly from the radiance field without tone mapping [2111.14451].

HDR DNVS becomes substantially harder when the observations are not clean, sharp, exposure-bracketed stills. In the single-exposure blurry-image-and-event setting of See-NeRF, blurry LDR images are modeled as exposure integration followed by a nonlinear camera response,
$$
\mathcal{I}_{LDR} = f\left(\int_{t_{start}}^{t_{end}} L(t)\,dt\right),
$$
which explicitly couples motion blur and dynamic range compression [2601.15475]. The same work treats event generation as operating on temporal and spatial derivatives of log-radiance, namely $\partial_t \log L(x,y,t)$ and $\nabla_{xy} \log L(x,y,t)$, and uses this complementarity to recover sharp HDR content lost to blur and clipping [2601.15475].

Several papers frame the central difficulty as an ill-posed inversion. Pano-NeRF states that HDR content is not recoverable in saturated regions of LDR inputs, and addresses this by transferring constraints from unsaturated regions through irradiance-radiance attenuation [2312.15942]. HDR-HexPlane identifies a second ill-posedness: if exposure and camera response are both unknown, their effects can compensate for each other, making the system underconstrained [2401.06052]. GaussHDR formulates a third ambiguity internal to the rendering pipeline, namely the tradeoff between 3D tone mapping and 2D tone mapping: the former yields good LDR fitting but unstable HDR predictions, while the latter stabilizes HDR fitting but can degrade LDR rendering fidelity [2503.10143].

This suggests that HDR DNVS is not a single inverse problem but a family of coupled inverse problems involving radiance recovery, exposure estimation, camera response calibration, temporal coherence, blur inversion, and sensor-specific measurement models.

## 2. Radiance-field lineage: from HDR-NeRF to dynamic HDR fields

HDR-NeRF establishes the canonical NeRF-based formulation for HDR view synthesis from LDR supervision. It recovers an HDR radiance field from LDR views with different exposures, generates both novel HDR views and novel LDR views under different exposures, and learns the camera response implicitly through three separate tone-mapping MLPs for R, G, and B [2111.14451]. Training is supervised only with LDR images taken at known camera poses and known exposure durations, and the paper introduces a unit exposure loss to reduce the scale ambiguity between radiance and exposure [2111.14451].

Pano-NeRF extends the radiance-field perspective by arguing that, in panoramic images, radiance from each pixel can be modeled as both a signal to convey scene lighting information and a light source to illuminate other pixels. It therefore introduces irradiance fields from sparse LDR panoramic images and jointly optimizes radiance and irradiance fields in a modified Mip-NeRF architecture that also predicts albedo [2312.15942]. The outgoing irradiance-driven radiance is expressed through an integral over incident directions modulated by BRDF and the cosine term, and the method uses a geometry regularization promoting thin and smooth surfaces together with an albedo chromaticity prior [2312.15942]. In evaluation, Pano-NeRF reports geometry metrics such as RMSE for depth and MAE for normals, as well as HDR metrics including PU-PSNR, PU-SSIM, HDR-VDP3, and RMSE [2312.15942].

HDR-HexPlane introduces a dynamic HDR NeRF framework built on HexPlane for dynamic scenes captured with various exposures. Its architecture explicitly separates dynamic scene representation and exposure modeling: HexPlane predicts log radiance and density, a learnable exposure mapping function estimates per-image exposure values, and a fixed monotonic camera response function stabilizes optimization when exposures are unknown [2401.06052]. The paper uses the sigmoid function as the CRF,
$$
g(x)=\frac{1}{1+e^{-x}},
$$
precisely because a learnable CRF combined with unknown exposures is unstable [2401.06052]. The method further constructs a dataset containing multiple dynamic scenes captured with diverse exposures and reports that it can render high-quality novel-view images at any time point with any desired exposure [2401.06052].

A plausible implication is that the radiance-field lineage splits along two axes. One axis concerns what is represented—radiance alone, or radiance plus irradiance, reflectance, illumination, or sensor response. The other concerns what is known at training time—exposure, CRF, geometry, or calibration—which directly determines whether the learning problem is well-conditioned.

## 3. Temporal coherence, events, and sensor-physics grounding

HDR-4DGS recasts the field in explicitly dynamic terms by introducing the task of HDR DNVS and proposing a 4D Gaussian Splatting architecture with a dynamic tone mapping module [2509.21853]. The 4D scene representation extends Gaussian splatting to spatiotemporal Gaussians, and the rendered image at time $t$ depends on temporal Gaussian weights $p_i(t)$ and spatial Gaussians $p_i(u,v|t)$ [2509.21853]. The dynamic tone mapping module is built from a radiance bank, a dynamic radiance context learner, and an exposure-aware MLP that maps HDR color to LDR color using both $\log(\mathbf{c}_t^h)+\log e_t$ and a temporal context embedding $\mathbf{f}_t$ [2509.21853]. The paper reports that static tone mapping or standard MLP tone mappers yield lower PSNR and higher LPIPS than context-aware mapping, and identifies a temporal window size of $k=20$ as the best tradeoff of accuracy and efficiency [2509.21853].

See-NeRF addresses a different dynamic failure mode: single-exposure blurry LDR images in extreme lighting conditions. Its central claim is that previous event-augmented HDR NeRF approaches ignore the sensor-physics misalignment between the rendered scene radiance and the actual measurement processes of RGB and event sensors [2601.15475]. See-NeRF therefore introduces two differentiable sensor-aware mapping fields. The pixel-wise RGB mapping field models exposure integration and nonlinear CRF by discretizing the exposure window, integrating HDR radiance over time, and mapping the integrated signal through a learned per-channel CRF MLP. The event mapping field models Bayer pattern adaptation, temporal delay compensation through a learned MLP and a first-order low-pass filter, and photometric quantity calibration via an estimated offset [2601.15475]. These mapping fields are jointly optimized with the NeRF network using an LDR image MSE and an event-count MSE:
$$
\mathcal{L}=\lambda \mathcal{L}_{evs}+\mathcal{L}_{ldr}.
$$
The reported outcome is state-of-the-art deblurring HDR novel view synthesis from single-exposure blurry LDR images and corresponding events [2601.15475].

A third temporal direction appears in low-light heterogeneous capture. “Learning Novel View Synthesis from Heterogeneous Low-light Captures” decomposes each image into reflectance, illumination, and noise under a robust Retinex model,
$$
S = R \odot L + N,
$$
with reflectance treated as multi-view invariant, illumination modeled by a per-view embedding and a learned tone-mapping module, and noise represented by a distinct trainable 2D map for each view [2403.13337]. The paper additionally introduces an illumination adjustment module that supports brightening and darkening through a control ratio $\epsilon$, trained with same-view images captured under higher and lower exposure [2403.13337].

Taken together, these works show that temporal coherence in HDR DNVS is not only a geometry-and-motion problem. It is also a tone-mapping problem, an exposure problem, and, in event-based settings, a sensor transfer problem.

## 4. Gaussian splatting and the shift toward explicit, efficient HDR representations

The migration from NeRF-style volumetric rendering to Gaussian splatting is one of the defining developments in HDR novel view synthesis. HDR-GS introduces a Dual Dynamic Range Gaussian point cloud model in which each Gaussian stores HDR color, LDR color, opacity, covariance, and tone-mapper parameters [2405.15125]. HDR color is modeled with spherical harmonics in log-exponential form, while LDR color is obtained through an MLP-based tone mapper applied to $\log \mathbf{c}^h + \log \Delta t$ [2405.15125]. The framework uses two Parallel Differentiable Rasterization processes to reconstruct HDR and LDR views and reports that it surpasses HDR-NeRF by 3.84 dB on LDR and 1.91 dB on HDR while achieving 1000x inference speed and requiring 6.3% training time [2405.15125].

GaussHDR modifies the Gaussian-splatting pipeline by unifying 3D and 2D local tone mapping through residual local tone mappers conditioned on context features [2503.10143]. It rasterizes per-Gaussian context features to the image plane, performs both 3D local tone mapping and 2D local tone mapping, supervises both LDR renderings, and adaptively modulates them with learned uncertainties [2503.10143]. The paper’s argument is that global tone mapping is too restrictive and that the balance between 3D and 2D tone mapping is scene dependent [2503.10143].

PhysHDR-GS reframes the appearance model itself. Rather than treating HDR content as a single radiance quantity to be inferred through tone-mapped supervision, it decomposes outgoing radiance into intrinsic reflectance and adjustable ambient illumination,
$$
L_o(\mathbf{x},\omega_o)=L_e(\mathbf{x})+L_a(\mathbf{x})\,H_r(\mathbf{x},\omega_o),
$$
and introduces an image-exposure branch together with a Gaussian-illumination branch [2603.28020]. The cross-branch HDR consistency loss
$$
\mathcal{L}_{cons} = \|\mathcal{G}(I_{HDR}\times t)-\mathcal{G}(\hat{I}_{HDR})\|_1
$$
provides explicit supervision for HDR content, while illumination-guided gradient scaling mitigates exposure-biased gradient starvation in under- and over-exposed regions [2603.28020]. The paper reports a PSNR gain of 2.04 dB over HDR-GS and rendering speed of up to 76 FPS [2603.28020].

LE3D focuses on RAW nighttime scenes and replaces spherical harmonics with a Color MLP and a primitive-aware color bias because SH is said to be unsuitable for RAW linear color space [2406.06216]. It adds Cone Scatter Initialization to compensate for poor SfM under extremely low SNR, and introduces depth distortion and near-far regularizations to improve structure for refocusing and other downstream tasks [2406.06216]. Compared with RawNeRF, LE3D reduces training time to 1% and improves rendering speed by up to 4,000 times for 2K resolution images in terms of FPS [2406.06216].

InstantHDR extends Gaussian splatting in a different direction: generalizable feed-forward inference. It reconstructs 3D HDR scenes from uncalibrated multi-exposure LDR collections in a single forward pass, combining a frozen geometry branch, geometry-guided appearance modeling for multi-exposure fusion, and a meta-network that predicts scene-specific tone-mapper parameters [2603.11298]. The paper introduces HDR-Pretrain, a dataset of 168 Blender-rendered scenes with diverse lighting types and multiple camera response functions, and reports $\sim700\times$ and $\sim20\times$ reconstruction speed improvement in single-forward and post-optimization settings [2603.11298].

| Method | Inputs | Distinguishing mechanism |
|---|---|---|
| HDR-GS | Multi-exposure LDR views | Dual Dynamic Range Gaussian model and Parallel Differentiable Rasterization |
| GaussHDR | Multi-view LDR images captured at different exposure levels | Unified 3D and 2D local tone mapping with uncertainty learning |
| PhysHDR-GS | Multi-exposure LDR views | Intrinsic reflectance plus adjustable ambient illumination with cross-branch HDR consistency |
| LE3D | RAW images, especially nighttime scenes | Cone Scatter Initialization, Color MLP, depth distortion and near-far regularizations |
| InstantHDR | Uncalibrated multi-exposure LDR collections | Feed-forward geometry-guided fusion and meta-network tone mapping |

A common misconception is that Gaussian splatting solves HDR NVS merely by accelerating rendering. The Gaussian-splatting literature instead alters representation, tone mapping, supervision, and optimization behavior, particularly in exposure-extreme regions [2405.15125] [2503.10143] [2603.28020].

## 5. Single-exposure, sparse, and low-light regimes

Multiple-exposure capture remains a strong assumption in much of HDR NVS, but several papers explicitly target the single-exposure regime. Mono-HDR-3D introduces, for the first time, the single-exposure HDR-NVS problem and proposes two modules based on LDR image formation principles: one converts LDR colors to HDR counterparts, and the other transforms HDR images back to LDR so that unsupervised learning is enabled in a closed loop [2505.01212]. The paper formulates the forward image model as
$$
I_l = \begin{cases}
\frac{\Delta t}{g}I_h + I_0 + \epsilon, & \text{Unsaturated} \\
I_{\max}, & \text{Saturated},
\end{cases}
$$
and implements an LDR-to-HDR color converter together with an HDR-to-LDR color converter that can be attached to NeRF or 3DGS [2505.01212]. In the reported results, Mono-HDR-GS achieves 41.68 LDR-PSNR and 38.57 HDR-PSNR, while HDR-GS and HDR-NeRF degrade substantially when forced to learn from only single-exposure data [2505.01212].

SeHDR approaches the same problem through 3D Gaussian bracketing. It begins with base 3D Gaussians learned from single-exposure multi-view LDR inputs, estimates multiple virtual Gaussians with identical geometry but varying linear colors conditioned on exposure manipulations, and then fuses them using Differentiable Neural Exposure Fusion (NeEF) [2509.20400]. The exposure manipulation is performed in linear radiance space through
$$
c_i^{l_j} = c_i^{l_1}\cdot \frac{\Delta t_j}{\Delta t_1},
$$
and the fusion weights are based on a learned well-exposedness score after a fixed CRF [2509.20400]. On the HDR-NeRF benchmark, SeHDR without HDR supervision reports 25.37 PSNR, 0.81 SSIM, and 0.117 LPIPS, and the paper states that it outperforms all competing methods without ground-truth HDR by wide margins [2509.20400].

Single-exposure assumptions also arise in more specialized sensing conditions. See-NeRF uses single-exposure blurry LDR images plus events rather than bracketed stills [2601.15475]. LE3D operates on noisy single-exposure RAW inputs, particularly for nighttime scenes, and supports exposure and tone-mapping changes at runtime because the representation is in RAW linear space [2406.06216]. Pano-NeRF, although not formulated as a single-exposure method, is explicitly designed for sparse LDR panoramic images and addresses HDR reconstruction under sparse-input constraints via irradiance fields [2312.15942].

A plausible implication is that the single-exposure literature has bifurcated into two strategies. One strategy injects additional physics or auxiliary sensing—events, RAW sensor modeling, or explicit forward–inverse imaging loops. The other synthesizes a virtual exposure stack in representation space, as in 3D Gaussian bracketing.

## 6. Evaluation, perceptual assessment, and recurring technical debates

Evaluation in HDR DNVS is heterogeneous because the field simultaneously cares about radiometric fidelity, perceptual appearance, temporal stability, and rendering efficiency. Across the cited works, common image metrics include PSNR, SSIM, and LPIPS; HDR-specific metrics include PU-PSNR, PU-SSIM, HDR-VDP3, and RMSE; and system metrics include FPS and training time [2312.15942] [2509.21853] [2405.15125]. Dynamic datasets and benchmarks now include the HDR-HexPlane dynamic dataset with multiple dynamic scenes captured with diverse exposures, HDR-4D-Syn and HDR-4D-Real for dynamic HDR evaluation, and HDR-Pretrain for generalizable feed-forward training [2401.06052] [2509.21853] [2603.11298].

Perceptual assessment is itself an active topic. “Perceptual Assessment and Optimization of HDR Image Rendering” proposes a family of HDR quality metrics that decomposes an HDR image into a stack of LDR images using an inverse display model,
$$
L^{(k)}=\left(\left[\frac{H\cdot v^{(k)}-b}{1-b}\right]_0^1\right)^{1/\gamma},
$$
then evaluates the decomposed images with existing LDR metrics and aggregates them with exposure- and spatial-weighting [2310.12877]. The paper emphasizes three properties: direct inheritance of advances in LDR quality metrics, no need for retraining on human HDR perceptual data, and the ability to prioritize specific luminance ranges [2310.12877]. It also reports perceptual optimization gains when used as a loss for HDR-NeRF† on synthetic DNVS scenes [2310.12877].

Several technical debates recur across the literature. One concerns whether tone-mapped supervision is sufficient for HDR reconstruction. PhysHDR-GS explicitly argues that constraining tone-mapped results fails in correcting abnormal HDR values and produces limited gradients in under- and over-exposed regions [2603.28020]. Another concerns whether tone mapping should be global, local, static, or dynamic. GaussHDR criticizes global tone mappers for impeding both HDR and LDR learning, while HDR-4DGS argues that static tone mapping fails to maintain temporal radiance coherence in dynamic scenes [2503.10143] [2509.21853]. A third concerns calibration and identifiability: HDR-HexPlane fixes the CRF to make exposure learning well-conditioned when exposures are unknown, whereas HDR-NeRF learns the CRF when exposures are known [2401.06052] [2111.14451].

The field therefore presents a layered picture rather than a single consensus pipeline. Some methods prioritize physical image formation and sensor transfer, some prioritize efficient explicit representations, some prioritize dynamic tone mapping and temporal coherence, and some prioritize deployability under single-exposure, low-light, or uncalibrated capture. What unifies them is the premise that HDR DNVS requires an explicit account of how scene radiance is measured, compressed, perturbed, and re-rendered across viewpoint and time.

Source: https://www.emergentmind.com/topics/hdr-dynamic-novel-view-synthesis-hdr-dnvs