Physically Inspired Gaussian Splatting for HDR Novel View Synthesis

Abstract

High dynamic range novel view synthesis (HDR-NVS) reconstructs scenes with dynamic details by fusing multi-exposure low dynamic range (LDR) views, yet it struggles to capture ambient illumination-dependent appearance. Implicitly supervising HDR content by constraining tone-mapped results fails in correcting abnormal HDR values, and results in limited gradients for Gaussians in under/over-exposed regions. To this end, we introduce PhysHDR-GS, a physically inspired HDR-NVS framework that models scene appearance via intrinsic reflectance and adjustable ambient illumination. PhysHDR-GS employs a complementary image-exposure (IE) branch and Gaussian-illumination (GI) branch to faithfully reproduce standard camera observations and capture illumination-dependent appearance changes, respectively. During training, the proposed cross-branch HDR consistency loss provides explicit supervision for HDR content, while an illumination-guided gradient scaling strategy mitigates exposure-biased gradient starvation and reduces under-densified representations. Experimental results across realistic and synthetic datasets demonstrate our superiority in reconstructing HDR details (e.g., a PSNR gain of 2.04 dB over HDR-GS), while maintaining real-time rendering speed (up to 76 FPS). Code and models are available at https://huimin-zeng.github.io/PhysHDR-GS/.

Figures (9)

• Figure 1: Variation of camera exposure $\Delta t$ and ambient illumination $\Delta L_a$ scales the HDR signal in different ways: $\Delta t$ causes a global change $\Delta I_{HDR}$, while $\Delta L_a$ induces local changes $\Delta \hat{I}_{HDR}$ (e.g., nameplate of the luckycat) via lighting-conditioned radiance variation $\Delta L_o$. Their different response patterns reveal complementary ways of modeling dynamic-range details.
• Figure 2: Overview of the proposed PhysHDR-GS, where Gaussian color is modeled from intrinsic reflectance and ambient illumination. The image–exposure (IE) branch modulates exposure $t$ on 2D images, while the Gaussian–illumination (GI) branch modulates ambient illumination $L_a$ on 3D Gaussians, yielding complementary dynamic-range details. Tone mapper $f$ performs tone mapping and dual-branch fusion for final LDR results. During training, a cross-branch HDR consistency loss $\mathcal{L}_{cons}$ enables explicit HDR self-supervision. Illumination-Guided Gradient Scaling rescales per-Gaussian gradients with $s_a$ to mitigate under-splitting in extreme exposure regions.
• Figure 3: Illustration of the tone mapper $f$. Given inputs $I_{HDR}\times t$ and $\hat{I}_{HDR}$, the tone-mapping MLP $f_{tm}$ first predicts global and local LDR outputs. The fusion MLP $f_{mix}$ then cross-fuses these global-local pairs to produce the final LDR result $I_{LDR}$.
• Figure 4: Gradient and illumination deviation analysis, where over/under-exposed pixels lie in flat regions of the tone mapping curve and yield a small Gaussian gradient. The gradient shows positive correlation with reciprocal illumination deviation $1/\Delta L_a$.
• Figure 5: Qualitative comparisons on LDR views. For each method, we show the reconstructed LDR image and the residual map w.r.t. the ground truth. Competing methods exhibit noticeable missing content in saturated regions (e.g., screen reflections in the 1st row), indicating information loss after tone mapping, whereas our method effectively preserves fine structures and details.