acceptodds
Under review as a conference paper at ICLR 2027

Prior-Guided Single-Exposure HDR Novel View Synthesis via Base-Gain Representation

Abstract

Single-exposure high dynamic range (HDR) novel view synthesis (NVS) recovers HDR scenes from low dynamic range (LDR) views captured at a single exposure and estimates camera response function (CRF) for LDR re-exposure. Without HDR ground-truth, using single-exposure LDR views as the sole source of direct supervision underconstrains HDR reconstruction and cannot ensure its faithful recovery, motivating the need for learned LDR-to-HDR reconstruction priors. Moreover, single-exposure LDR views supervise CRF only over a limited range, necessitating a learnable prior for reliable re-exposure beyond it. We propose PriorGS, a 3D Gaussian Splatting framework that addresses both challenges by (1) guiding HDR recovery with pseudo-HDR supervision from 2D single-exposure HDR reconstruction methods while anchoring it to input LDR views through a closed reconstruction loop; and (2) regularizing CRF modeling with a learnable prior that guarantees monotonicity and promotes plausible responses. Yet cross-view inconsistencies in independently generated pseudo-HDR targets impair scene reconstruction when HDR is encoded directly as Gaussian's color attribute. We therefore introduce a 3D base–gain HDR representation that factors HDR radiance into an LDR base and an overlaid 3D gain field, promoting robustness to such inconsistencies and enabling 3D HDR appearance editing. Extensive experiments across public datasets demonstrate PriorGS's substantial superiority over baselines. A short demo video is included in the supplementary materials.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
HDR novel view synthesis, single-exposure HDR reconstruction, 3D Gaussian Splatting, HDR Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5TVdOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel method for 3D HDR novel view synthesis, specifically achieved through pseudo-supervision via a 2D HDR reconstruction approach, a Base-Gain HDR representation, and Prior-Constrained CRF modeling.

Critical strengths
  1. It seems to have a very good effect.
  2. The motivation for constructing the entire module is clear.
Critical weaknesses
  1. The implementation appears somewhat engineering-oriented, and the current approaches are quite common, with no particularly novel findings. For example, 3D base-gain extends the 2D approach to 3D, and closed-loop design, which already exists in Mono-HDR-3D, is also incorporated.
  2. The performance of the HDR-supervised model drops significantly without supervision. This raises the question of whether all observed gains stem from the pretrained model. If a model with limited generalization capability were used instead, would performance decline substantially? It is worth noting that datasets of a similar scale are typically quite small.
Questions for authors

1.The substantial gain is largely attributed to the pretrained HDR supervision. If this HDR model were replaced with one not trained on this domain, how would the model’s performance change? Is the current comparison fair, given that this pretrained prior has been utilized? Moreover, this is not a widely used prior in the conventional sense. 2. What is the innovative contribution of the Base-Gain representation? It appears to be a straightforward extension of a conventional gain diagram. I do not seem to have a thorough understanding of its core novelty. 3.The monotonic CRF approach had already appeared by 2017. The authors' contribution relative to prior work lies not in merely applying a similar method to 3D or other domains—such straightforward application is largely an engineering task—but rather in the specific innovations they introduce.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5TVd (1/4)

We thank Reviewer 5TVd for the constructive comments and for recognizing the clear motivation and strong performance of our method. We hope our responses could address all the concerns.

W2 & Q1: All gains may stem from the pretrained HDR prior

In short: based on our results, the HDR prior is a key ingredient but not the sole source of our gains; even without it, PriorGS outperforms all baselines on most metrics (Table 2(a) vs. Table 1; main paper). Also, given the same HDR prior, a baseline and our ablations that remove or replace our design (Table 2; main paper) all trail PriorGS. Details follow.

Thank you for this question. We agree that removing the HDR prior causes a large drop (Table 2(a)); this is consistent with our finding that single-exposure HDR-NVS is underconstrained without such a prior (Figs. 1, 4). However, three pieces of experimental evidence show that our gains do not stem from the HDR prior alone:

(1) Performance w/o the prior (Table 2). Even after prior drop, PriorGS (Table 2(a)) outperforms all baselines (Table 1) on HDR-NeRF-Real in every metric and on HDR-Plenoxels-Real in SSIM and LPIPS.

(2) New prior-matched baseline. We additionally trained HDR-GS with the same prior (IntrinsicHDR pseudo-HDR targets) as PriorGS, all else unchanged. We chose HDR-GS as it has the strongest HDR recovery in Table 1 (main paper) among baselines that natively support HDR supervision. With the HDR prior now shared, the two differ in our base–gain representation, closed-loop reconstruction, and prior-constrained CRF. PriorGS still significantly outperforms HDR-GS + Prior in all comparisons:

HDR-NeRF-Real – LDR│HDR-Plenoxels-Real – LDR
MethodPSNR ↑SSIM ↑LPIPS ↓│PSNR ↑SSIM ↑LPIPS ↓
HDR-GS11.100.52120.370│4.700.27990.456
HDR-GS + Prior (IntrinsicHDR)17.860.77900.193│12.970.63740.254
PriorGS (Ours)24.310.92480.064│23.000.90210.092
HDR-NeRF-Syn – HDR│HDR-NeRF-Syn – LDR
MethodPSNR ↑SSIM ↑LPIPS ↓│PSNR ↑SSIM ↑LPIPS ↓
HDR-GS26.110.89390.113│10.860.34290.397
HDR-GS + Prior (IntrinsicHDR)30.670.94120.059│13.730.58280.329
PriorGS (Ours)31.220.95690.038│22.960.89370.081

We will add this new comparison to the paper.

(3) Existing ablations (Table 2). Every variant that keeps the prior but removes or replaces one of our components (Table 2(b), (c), (f–h)) is worse on all six metrics.


Further results. In part 4 of our response to Reviewer EB8P (W3 & Q4), a stronger same-prior baseline also trails PriorGS on every metric across all datasets and tasks.

AuthorsOpenReview ↗

Response to Reviewer 5TVd (2/4)

W2 & Q1: Performance with a weaker or out-of-domain HDR prior

In short: weaker HDR priors do not erase our superiority; with the four priors we test (Table 3, Appendix A.4), PriorGS beats the strongest baseline in 47 of 48 comparisons. All four priors, IntrinsicHDR included, are already out-of-domain: none was trained or fine-tuned on our benchmarking datasets, and since each was trained on different data, our advantage does not stem from a domain-matched prior. Details follow.

Thank you for raising this point.

  • Weaker priors. Appendix A.4 (Table 3) replaces IntrinsicHDR with three other single-image HDR models: ExpandNet (2018), DeepHDR (2020), and GMNet (2025). With every prior, PriorGS outperforms the strongest baseline on all 12 metrics of Table 3, with a single exception: PSNR on HDR-Plenoxels-Real with GMNet (16.9 vs. 17.7 dB), where PriorGS still wins in SSIM and LPIPS. Prior quality matters mainly on real scenes (e.g., 19.4–24.3 dB PSNR on HDR-NeRF-Real across priors, all above the best baseline’s 16.3 dB), whereas on synthetic scenes all four priors lie within 1.4 dB HDR-PSNR of each other. Consequently, PriorGS’s superiority is not tied to IntrinsicHDR, and our framework can directly benefit from other existing or stronger future HDR reconstruction models.

  • Out-of-domain priors. We use each model’s public weights without fine-tuning, and no benchmarking dataset, including the HDR GT of synthetic scenes (L298–299), is involved in training any prior. ExpandNet and DeepHDR were even released before either benchmark existed. Moreover, each prior was trained on a different selection of scenes and capture settings: synthetic indoor renders and multi-illumination photos (IntrinsicHDR), HDRTV videos and mobile-phone captures (GMNet), and two separately collected sets of HDR photos and videos (ExpandNet, DeepHDR). Although HDR datasets are small, as the reviewer notes, priors trained on such different data all keep PriorGS ahead of the baselines on scenes they never saw; hence, our advantage does not stem from a prior trained on a similar domain.

Q1: Fairness of the comparison + not a widely used prior

In short: we believe the comparison is fair, as all methods receive the same single-exposure LDR views and no HDR GT; our prior only adds zero-shot knowledge from a pretrained 2D model, as is common practice in 3D reconstruction. Details follow.

Thank you for this question.

  • Fairness. All methods, including PriorGS, train on the same single-exposure LDR views without HDR GT (L298–299). The frozen 2D model processes only these views, so it adds learned knowledge rather than additional observations of the scene. Likewise, methods using pretrained depth/normal priors in sparse-view NVS and surface reconstruction (e.g., MonoSDF [1], FSGS [2], DNGaussian [3]) are routinely compared against baselines without such priors.
  • Not a widely used prior. Supervising a 3D representation with a pretrained 2D model is standard practice [1–3]; our prior is its HDR counterpart. HDR priors are less common than depth/normal priors because HDR reconstruction is a more specialized task.

[1] Yu, Zehao, et al. "Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction." NeurIPS 2022.

[2] Zhu, Zehao, et al. "Fsgs: Real-time few-shot view synthesis using gaussian splatting." ECCV 2024.

[3] Li, Jiahe, et al. "Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization." CVPR 2024.

AuthorsOpenReview ↗

Response to Reviewer 5TVd (3/4)

W1.1 Our novel findings

Thank you for raising these points, which help us clarify our novelty. PriorGS is built on three novel findings about single-exposure HDR-NVS:

  • Novel Finding 1: Using single-exposure LDR views as the only source of direct supervision leaves HDR reconstruction underconstrained (L050-052).

    This motivates the need for learned LDR-to-HDR reconstruction priors to guide HDR recovery (first contribution). Accordingly, prior-free methods fail at recovering HDR radiance although they reproduce the LDR views faithfully (Fig. 1).

  • Novel Finding 2: Anchoring the scene to reliable LDR views while allocating dedicated capacity to less reliable HDR extension improves robustness to cross-view pseudo-HDR inconsistencies (L101-106).

    Our base-gain representation (second contribution) realizes this with modeling HDR radiance as an LDR base supervised by LDR views and extended to HDR through a 3D gain field. In contrast, direct HDR encoding of equal capacity suffers severe color and geometry artifacts (Table 2(c); Figs. 2, 7, 16). The 3D gain field also enables 3D HDR appearance editing (Figs. 6, 13, 14).

  • Novel Finding 3: CRF modeling in single-exposure HDR-NVS faces a fundamental dilemma: the CRF must be learned per scene from a single exposure, yet generalize to unseen exposures beyond the narrow supervised range (L077-081).

    Existing methods sacrifice one side or the other: flexible CRFs adapt but overfit the observed exposure, while fixed CRFs avoid overfitting but lack scene adaptability (Table 1 vs. 4; Figs. 5, 10). We address this dilemma by regularizing CRF modeling (third contribution) with a learnable prior that constrains the response beyond the supervised range while adapting to each scene, consistently outperforming flexible and fixed alternatives (Table 2(f)–(h); Figs. 11, 18).

W1.2 Research significance of PriorGS

While PriorGS builds on established components, its research significance lies in what it reveals about single-exposure HDR-NVS. Our novel findings above contribute knowledge in two ways:

  • they diagnose why existing approaches fail; and
  • they show how these failures can be addressed, from which our implementation is derived and is therefore knowledge-driven rather than engineering-oriented.

This knowledge is not specific to PriorGS or even to single-exposure HDR-NVS: it extends to 3D reconstruction from degraded observations guided by learned 2D priors, where the same challenges of underconstrained recovery and inconsistent per-view priors arise.

W1.3 How our closed-loop design differs from Mono-HDR-3D’s

Thank you for pointing this out; we will update the related works section to better present the distinction. As our contribution list reflects:

We do not claim closed-loop reconstruction as a standalone novelty: our novelty lies in incorporating learned LDR-to-HDR priors into HDR recovery, with the closed loop designed specifically to make this incorporation more reliable.

Although structurally similar, the two loops differ in motivation and role:

  • Mono-HDR-3D (Mono-HDR-GS in our paper) uses its loop as a substitute for the HDR supervision it lacks, yet the loop alone leaves HDR recovery underconstrained (Figs. 1, 4).
  • PriorGS uses its loop as a complement to imperfect pseudo-HDR supervision, anchoring the prior-guided HDR to the reliable LDR views at both ends.

Our ablations reflect this distinction: the loop without the prior falls far short (Table 2(a)), while adding the loop to the prior further improves reconstruction (Table 2(b); Figs. 7, 15).

Reviewer_EB8POpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PriorGS does HDR novel-view synthesis and LDR re-exposure from posed single-exposure images, without scene HDR ground truth. A pretrained single-image HDR model gives pseudo-HDR targets. Each Gaussian has a supervised LDR base and a view-dependent log-gain on shared geometry. A learnable four-parameter monotonic CRF maps renderings back to the input exposure, closing a reconstruction loop. The gain field also allows global and local editing. Experiments use HDR-NeRF (synthetic and real) and HDR-Plenoxels, with one model per input exposure tested on the others; synthetic scenes add HDR ground truth. Ablations, other pseudo-HDR generators and a stronger Scaffold-GS GaussHDR baseline are included.

Critical strengths
  1. With one input exposure, the full system clearly beats the four baselines it compares against (HDR-GS, GaussHDR, Mono-HDR-GS and SeHDR) on HDR novel-view synthesis and re-exposure, including unseen exposures. It also beats the stronger Scaffold-GS GaussHDR (Appendix A.4, Table 6): 31.22 vs. 28.45 dB synthetic HDR PSNR, and 24.31 vs. 22.76 dB (HDR-NeRF) and 23.00 vs. 21.03 dB (HDR-Plenoxels) re-exposure PSNR, with better SSIM/LPIPS. Much of this gap comes from the 2D HDR prior, which the baselines lack (Weakness 3). Table 2(c) suggests the design also matters: with the same prior, the full base-gain model beats direct HDR encoding (22.71 to 24.31 dB on HDR-NeRF-Real, 15.05 to 23.00 dB on HDR-Plenoxels), though this comparison may also change the supervision (Weakness 1).

  2. The design is coherent: pseudo-HDR targets supply the missing radiance, LDR supervision anchors the scene, and a compact monotonic CRF, whose four-parameter form is checked against 201 DoRF responses (Appendix A.2), regularizes re-exposure. The per-Gaussian log-gain field, lifted from 2D gain maps, is the new piece, and it also enables global and local editing. The other parts are known (per-view 2D pseudo-HDR supervision, a re-exposure loss, a DoRF-based CRF), so this is solid system work rather than a new problem formulation.

Critical weaknesses
  1. The multiplicative base-gain coupling is not separated from the extra LDR supervision. Table 2(c) and Appendix A.7 compare with direct log-HDR encoding at a higher SH degree, and Figure 19 shows qualitatively that dropping LDR-base supervision hurts. But the direct variant (one SH color set for log-radiance) seems to lack the LDR head and L_base. If so, (c) changes representation and supervision at once, and Figure 19 only shows that L_base matters within the base-gain model, not whether the factorization helps once LDR supervision is matched. A two-head control with matched supervision would settle this (Question 1). Nor is the specific per-Gaussian SH gain clearly better than an MLP gain: MLP(c,mu,d) reaches 24.44 vs. 24.31 dB on HDR-NeRF-Real and 22.64 vs. 23.00 dB on HDR-Plenoxels (Table 2(d)), although the paper reports texture artifacts for the MLP variants (Figure 17).

  2. Table 2 scores re-exposed LDR images on real scenes, mixing radiance and CRF errors. Appendix A.7 is only qualitative, and no component ablation is scored quantitatively against HDR ground truth (Table 3 only swaps the 2D generator). So the evidence backs the full system more than each component's role in HDR recovery. For example, the reconstruction loop moves HDR-Plenoxels re-exposure PSNR only from 22.97 to 23.00 dB (SSIM/LPIPS improve) but adds 0.79 dB on HDR-NeRF-Real (23.52 to 24.31 dB). Such dataset- and metric-dependent gaps are hard to read without HDR ground truth (Question 2).

  3. Table 1 (p. 6) does not separate the prior from the framework. Only PriorGS gets a pretrained single-image HDR model (IntrinsicHDR). Without it, PriorGS (18.39 / 15.00 dB, Table 2(a)) is in the same range as GaussHDR (16.28 / 17.73 dB): higher on HDR-NeRF-Real, lower on HDR-Plenoxels. So, at least for real-scene re-exposure, most of the "large margins" depend on the prior, and no such ablation exists for synthetic HDR; whether the framework adds anything once the baselines get the same prior is untested. Table 3 points the same way: on HDR-Plenoxels PriorGS ranges from 16.94 to 23.00 dB across 2D generators, and with GMNet it falls below GaussHDR's 17.73 dB. The simplest prior-equipped baseline, LDR 3DGS plus the same 2D HDR model per view, is also missing (Question 4). Some claims go too far. Section 4.2 calls the gain field a "unique mechanism" for this editing, but the only alternative tested (Figure 12) is linear scaling of HDR radiance, which is an exposure shift by construction. Log-domain scaling of a direct HDR encoding would also expand dynamic range. Some baseline numbers also seem off: the reproduced SeHDR is about 10 dB below its own paper, and HDR-GS at 4.70 dB on HDR-Plenoxels implies an RMS error of about 0.58, suggesting largely black or white outputs (see Question 5). This matters for Table 1: against SeHDR's own number, the synthetic HDR margin would be about 6 dB rather than 16 dB.

Minor points, to help the presentation and not part of my assessment: Figure 14 (p. 23) uses ρ = 0.25 both to brighten and to darken without saying the log-gain can be negative. Section 4.1 says five exposures per view; Appendix A.3 says two HDR-Plenoxels scenes have three. The dynamic-range metric in Figures 6/12 is not defined: Figure 12 gives 13.50 stops for HDR scaling at both ρ = 0.5 and 1.5, while Figure 6 gives 13.9 stops for the unscaled desk view, possibly because Figure 12 is restricted to the 1st–99th percentiles; please state how DR is computed.

Questions for authors
  1. Which color heads and losses remain in Table 2(c)? Is a separate LDR head with L_base kept? If not, could you point to or add a shared-geometry control with separate LDR and log-HDR heads and matched pseudo-HDR targets, reconstruction-loop loss and capacity? This would show whether the gain comes from the factorization or from the LDR supervision.

  2. With the same scene-wise correction, how do the prior, reconstruction-loop and representation ablations score on synthetic HDR ground truth, split by input-saturated, dark and well-exposed regions? This would show, with existing data, whether missing radiance is recovered or only re-exposure improves.

  3. Appendix A.3 fits the luminance-correction weights once on the first indexed view and reuses them for the scene. What is the correction function, how many parameters does it have, and what reference does that view give? Is that view included in the reported metrics? One correction per scene is fine; I just need the details to read the HDR scores.

  4. Could you add two same-prior baselines: (a) standard 3DGS LDR views lifted per view by IntrinsicHDR and re-exposed with the same CRF; (b) GaussHDR/HDR-GS trained on the same pseudo-HDR targets? This would split the Table 1 margin between the prior and the representation. It could change my rating.

  5. (a) SeHDR (ICCV 2025) reports 25.37 dB HDR PSNR on HDR-NeRF-Syn without HDR ground truth, under what seems to be the same protocol (one run per training exposure, 18/17 views, μ-law with μ = 5000). Your reproduction gets 15.14 dB (Table 1), even with your extra scene-wise luminance correction, while HDR-GS rises from 11.01 dB (SeHDR paper) to 26.11 dB under your protocol. SeHDR's code is not public: what did you reimplement, and how were its outputs in Tables 1, 3, 4 and 5 produced? Please also give per-exposure HDR-GS results on HDR-Plenoxels; 4.70 dB means an RMS error of about 0.58 on [0,1] intensities. (b) PhysHDR-GS (CVPR 2026), which you cite, also splits Gaussian appearance into intrinsic reflectance and an adjustable illumination term, with a Gaussian-level illumination branch. Its code is not released, so I do not expect experiments, but please briefly discuss how your base-gain representation and editing differ.

8 comments
AuthorsOpenReview ↗

Response to Reviewer EB8P (1/6)

We thank Reviewer EB8P for the constructive comments and for recognizing our novel and editable gain field, substantial improvements, and clear presentation. We hope our responses address your concerns.

W1 & Q1: Matched-supervision control and MLP variant

In short: We extended our ablations by adding the suggested two-head control with matched supervision and capacity. Even so, the base-gain factorization improves every metric on all three datasets, including against HDR GT. LDR supervision contributes too, explaining much of the deficit of Table 2(c). The MLP variant (Table 2(d)) trails PriorGS on five of six real-scene metrics and by 3.17 dB PSNR against HDR GT. Details follow.

Thank you for your feedback. We will add the following analysis to the paper.

Table 2(c) setup. (c) has a single SH color head per Gaussian encoding log-HDR radiance, with a higher SH degree to match our capacity (Appendix A.7). It keeps the pseudo-HDR and re-exposure losses but has no LDR head, hence no L_base. We agree that (c) thus changes both representation and supervision.

Matched-supervision control. Our new variant (c') has the same shared geometry, pseudo-HDR targets, losses (Eq. 7) and capacity as PriorGS, with separate LDR and log-HDR SH heads. It thus differs from PriorGS only in predicting log-HDR radiance freely rather than as a gain on the LDR base (Eq. 2).

(ID) DescriptionHDR-NeRF-Syn - HDRHDR-NeRF-Real - LDRHDR-Plenoxels-Real - LDR
PSNR ↑ / SSIM ↑ / LPIPS ↓PSNR ↑ / SSIM ↑ / LPIPS ↓PSNR ↑ / SSIM ↑ / LPIPS ↓
(c) Base-Gain → direct encoding28.52 / 0.8735 / 0.17022.71 / 0.8981 / 0.10515.05 / 0.6860 / 0.251
(c′) Base-Gain → direct enc. + LDR Sup.30.77 / 0.9399 / 0.06723.39 / 0.9238 / 0.07522.54 / 0.8974 / 0.097
PriorGS (Ours)31.22 / 0.9569 / 0.03824.31 / 0.9248 / 0.06423.00 / 0.9021 / 0.092

Even with matched supervision, base-gain wins on every metric. This confirms the contribution of the factorization itself: it lets L_base anchor HDR color through the base, not only the shared geometry. LDR supervision contributes too, recovering much of the gap between (c) and PriorGS.

Gain field vs. MLP. We agree that MLP(c, μ, d) is competitive on real-scene re-exposure. Still, PriorGS leads on five of six metrics there, trailing only by 0.13 dB PSNR on HDR-NeRF-Real (Table 2(d)), and avoids the MLP's texture artifacts in challenging regions (Figures 7 and 17). Against HDR GT on HDR-NeRF-Syn (our response to W2 & Q2), the MLP variant falls well short:

(ID) DescriptionPSNR ↑SSIM ↑LPIPS ↓
(d) Gain Field → MLP(c, μ, d)28.050.85360.126
PriorGS (Ours)31.220.95690.038
AuthorsOpenReview ↗

Response to Reviewer EB8P (2/6)

W2 & Q2: Additional quantitative ablations on HDR-NeRF synthetic data

In short: As requested, we extended our real-scene ablations to synthetic scenes with HDR GT and split the results by region. All ablations trail PriorGS overall and in saturated and well-exposed regions, consistent with our original ablations (Table 2) and qualitative HDR ablations; the prior recovers radiance lost to saturation (SSIM 0.92 vs. 0.70 without it). Details follow.

We thank the reviewer for this suggestion and will add the following analysis to the paper.

Requested analysis. We agree that HDR GT is the most direct measure and LDR re-exposure contains the CRF error. We ran ablations (a)-(e) on HDR-NeRF-Syn against HDR GT, with the same per-scene luminance correction as Table 1:

(ID) DescriptionPSNR ↑SSIM ↑LPIPS ↓
(a) Remove LDR-to-HDR prior29.070.92720.078
(b) Remove reconstruction loop31.010.95200.043
(c) Base-Gain → direct encoding28.520.87350.170
(d) Gain Field → MLP(c, μ, d)28.050.85360.126
(e) Gain Field → MLP(c)30.350.95640.039
PriorGS (Ours)31.220.95690.038

All ablations trail PriorGS on every HDR metric, consistent with the overall trends of our real-scene ablations (Table 2) and with our qualitative results (Figures 2, 7, 15-17 and 19).

For the region-wise analysis, we label each HDR pixel using the corresponding GT LDR image at the training exposure: dark if its luminance is below 0.1, saturated if above 0.9, and well-exposed otherwise. We report SSIM and LPIPS (averaged across each region) as they better reflect structural loss:

(ID) DescriptionSaturatedWell-exposedDark
SSIM ↑ / LPIPS ↓SSIM ↑ / LPIPS ↓SSIM ↑ / LPIPS ↓
(a) Remove LDR-to-HDR prior0.7042 / 0.25170.9427 / 0.06280.9530 / 0.0421
(b) Remove reconstruction loop0.9038 / 0.07050.9609 / 0.03720.9118 / 0.0497
(c) Base-Gain → direct encoding0.9009 / 0.06720.8874 / 0.15410.7969 / 0.2081
(d) Gain Field → MLP(c, μ, d)0.7868 / 0.17930.8633 / 0.11480.8216 / 0.1304
(e) Gain Field → MLP(c)0.9167 / 0.06050.9616 / 0.03390.9232 / 0.0469
PriorGS (Ours)0.9184 / 0.05930.9621 / 0.03350.9242 / 0.0452

All ablations trail PriorGS in saturated and well-exposed regions. The gap is largest in saturated regions without the prior: SSIM falls from 0.9184 to 0.7042, indicating that the prior recovers radiance lost to saturation. In dark regions, PriorGS ranks second, behind the prior-free variant (a). Unlike real captures, synthetic low-exposure views contain little shadow noise, so the input LDR already determines dark radiance with good precision, and the 2D prior mainly adds its own estimation error.

Why we ablated on real scenes. We ablated on real rather than synthetic scenes for their more challenging content and more diverse capture/exposure settings (Sec. 4.3). As they lack HDR GT, we assessed HDR recovery through re-exposure at unseen exposures, which expose the input's saturated (lower exposures) and dark (higher exposures) regions, an established way to evaluate HDR quality (Cao et al. [1]).

Regarding your remark:

The reconstruction loop moves HDR-Plenoxels re-exposure PSNR only from 22.97 to 23.00 dB (SSIM/LPIPS improve) but adds 0.79 dB on HDR-NeRF-Real (23.52 to 24.31 dB) ...

Please refer to our response to Q2 of reviewer 977U. Our analysis above complements our ablations in Table 2 and further confirms the contribution of closed-loop reconstruction.

[1] Cao et al. "Perceptual assessment and optimization of HDR image rendering." CVPR 2024.

AuthorsOpenReview ↗

Response to Reviewer EB8P (3/6)

Q3: Per-scene luminance correction details

In short: We apply Hanji et al.'s luminance correction (cited in Sec. 4.1) exactly as in their official code: a global cubic on PQ-encoded luminance with four coefficients. We fit it once per scene on the first test view against its HDR GT and reuse it for all test views. Details follow.

Thank you for confirming that one correction per scene is fine.

Correction function. We apply the luminance correction of Hanji et al. [1] (their Eq. 3) exactly as in their official MATLAB implementation: a global cubic on PQ-encoded luminance PP,

Pc=w3P3+w2P2+w1P+w0,P_c=w_3P^3+w_2P^2+w_1P+w_0,

followed by inverse PQ. As in their code, both images are first scaled to a median of 500, and the output is scaled back to the GT range. We keep the rendered chromaticity (omitting their Eq. 4), so color errors remain fully penalized.

Parameters and reference. For each method and scene, we fit w0w_0-w3w_3 by unregularized least squares over all pixels of the first test view, mapping its rendered luminance to its HDR GT luminance. We then fix the four coefficients and both median scales for all test views and include the fitting view in the reported metrics. Hanji et al. fit each test image to its own GT; we use a single GT view per scene, so our correction cannot mask flicker across views (Appendix A.3).

[1] Hanji et al. "Comparison of single image HDR reconstruction methods — the caveats of quality assessment." SIGGRAPH 2022.

Reviewer_977UOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PriorGS for single-exposure HDR novel view synthesis without HDR ground-truth supervision. The method uses pseudo-HDR targets from pretrained 2D HDR reconstruction models and introduces a 3D base–gain representation to mitigate cross-view inconsistencies. A prior-constrained learnable CRF is further incorporated to support LDR re-exposure through closed-loop reconstruction. Experiments on HDR-NeRF and HDR-Plenoxels report substantial improvements over the evaluated baselines in re-exposed LDR-NVS, as well as HDR-NVS on synthetic scenes. The paper also presents component-wise ablations and applications to 3D HDR appearance editing.

Critical strengths

(1) The proposed 3D base–gain representation provides a meaningful way to combine reliable LDR supervision with less reliable pseudo-HDR targets. By representing LDR appearance and HDR enhancement using shared Gaussian geometry, the method introduces a useful inductive bias for mitigating cross-view inconsistencies. The representation also enables view-consistent 3D HDR appearance editing.

(2) PriorGS reports substantial improvements over the evaluated baselines across real and synthetic scenes, particularly for HDR-NVS and re-exposure at unseen exposure levels. The paper also includes controlled comparisons of alternative HDR representations and CRF models, experiments with different 2D HDR reconstruction priors, and a mathematical analysis of the CRF formulation. These results provide meaningful evidence for the effectiveness of the overall design.

Critical weaknesses

(1) The novelty of some components could be more clearly distinguished from existing approaches. Although the proposed 3D base–gain representation offers a meaningful extension, several underlying ideas have precedents in prior work. Mono-HDR-3D (Zhang et al., 2025) has explored closed-loop HDR-to-LDR reconstruction, while SeHDR (Li et al., 2025) has identified cross-view inconsistencies arising from independently reconstructed 2D HDR targets. Moreover, LDR/SDR-base and gain-map decomposition have been explored in prior 2D HDR methods (Liao et al., 2025; Guan et al., 2025; Canham et al., 2025). More closely related, SeHDR already learns a base Gaussian representation and derives an HDR Gaussian representation with shared geometry. The paper should more clearly distinguish the proposed technical extensions from existing approaches, particularly in terms of their representational and methodological contributions.

(2) Some experimental comparisons and core methodological claims require further validation. First, the substantial discrepancy between the reproduced SeHDR results and those reported in the original publication raises concerns about baseline comparability. Second, the quantitative ablations primarily evaluate re-exposed LDR-NVS rather than HDR radiance recovery, making it difficult to fully assess the contributions of HDR radiance recovery. Moreover, the contribution of closed-loop reconstruction varies across datasets, while the robustness of the base–gain representation to cross-view inconsistencies in pseudo-HDR supervision has not been systematically evaluated. Further validation of the learned CRF would also strengthen the experimental evidence. Specific concerns are detailed in the Questions for Authors section.

Questions for authors

Q1. The reproduced SeHDR result in Table 1 differs substantially from that reported in the original SeHDR paper. In particular, SeHDR originally reports an HDR PSNR of 25.371 dB without HDR ground-truth supervision on the HDR-NeRF synthetic dataset, whereas the current paper reports 15.14 dB. Although differences in training or evaluation protocols may contribute to this discrepancy, the gap appears significant. A clarification of the reproduction procedure, including HDR normalization, luminance correction, and evaluation settings, would help establish the fairness of the comparison.

Q2. Although Table 2(b) evaluates the contribution of closed-loop reconstruction, the resulting improvements vary across datasets. In particular, the PSNR improvement on HDR-Plenoxels Real is only 0.03 dB (22.97 to 23.00 dB), compared with 0.79 dB on HDR-NeRF Real. A more detailed analysis of this difference would help clarify the contribution and necessity of the closed-loop design across datasets.

Q3. The quantitative ablations in Table 2 mainly evaluate re-exposed LDR-NVS on real scenes. Additional quantitative evaluations of the corresponding ablation variants on the HDR-NeRF synthetic dataset are required using HDR ground truth.

Q4. The base-gain representation is primarily motivated by the cross-view inconsistencies in independently generated pseudo-HDR targets. Although Table 2(c) and Figures 2 and 16 demonstrate improved reconstruction quality compared with direct HDR encoding, they do not directly evaluate the robustness of the proposed representation to varying degrees of cross-view inconsistency. Additional quantitative evaluations of HDR reconstruction quality and cross-view consistency under varying levels of inconsistency in pseudo-HDR targets would help clarify how effectively the proposed representation mitigates the adverse effects of inconsistent HDR supervision.

Q5. The proposed CRF is strictly monotonic by construction, and Appendix A.2 demonstrates its ability to approximate the DoRF response family. However, these results do not directly establish the accuracy of the scene-specific CRF recovered from single-exposure observations. A direct comparison between the learned CRFs and the ground-truth response curves on the HDR-NeRF synthetic dataset would strengthen this claim.

Q6. The base-gain representation introduces an additional set of view-dependent SH coefficients for each Gaussian, and the method also requires pseudo-HDR preprocessing. Reporting the resulting parameter overhead, rendering speed, memory consumption, and total preprocessing/training cost relative to existing HDR-NVS methods would help assess its practical efficiency.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 977U (1/5)

We thank Reviewer 977U for the constructive comments and for recognizing our meaningful base-gain representation, substantial improvements, extensive analyses, and clear presentation. We hope our responses address your concerns.

W1: Further distinguishing our novelties from existing approaches

Thank you for raising these points, which help us clarify our novelty. We will revise the paper to reflect the distinctions below.

Mono-HDR-3D has explored closed-loop HDR-to-LDR reconstruction.

As our contribution list reflects:

We do not claim closed-loop reconstruction as a standalone novelty: our novelty lies in incorporating learned LDR-to-HDR priors into HDR recovery, with the closed loop designed specifically to make this incorporation more reliable.

Although structurally similar, the two loops differ in motivation and role:

  • Mono-HDR-3D uses its loop as a substitute for the HDR supervision it lacks, yet the loop alone leaves HDR recovery underconstrained (Figs. 1, 4).
  • PriorGS uses its loop as a complement to imperfect pseudo-HDR supervision, anchoring the prior-guided HDR to the reliable LDR views at both ends.

Our ablations reflect this distinction: the loop without the prior falls far short (Table 2(a)), while adding the loop to the prior further improves reconstruction (Table 2(b); Figs. 7, 15; please also see our response to Q2).

SeHDR has identified cross-view inconsistencies arising from independently reconstructed 2D HDR targets.

Indeed, SeHDR makes this observation, and we will acknowledge it in the revision. The distinction lies in how each method responds to it:

  • SeHDR reports that training 3DGS on these targets is ineffective and thus forgoes 2D HDR priors, leaving its HDR recovery underconstrained.
  • We leverage the prior and propose a 3D base-gain representation that promotes robustness to its inconsistencies, which can impair scene reconstruction under direct HDR encoding (Table 2(c); Figs. 2, 16).

LDR/SDR-base and gain-map decomposition have been explored in prior 2D HDR methods (Liao et al., 2025; Guan et al., 2025; Canham et al., 2025).

Our paper credits these works as the inspiration for our design (Sec. 2 and 3.1). As our contribution list reflects:

The core novelty of our base-gain representation is improving robustness of 3D HDR reconstruction to cross-view inconsistencies in pseudo-HDR targets.

This 3D-specific novel functionality and its motivation are fundamentally different from those of 2D gain maps:

  • Motivation: Exploiting the HDR information in pseudo-HDR targets while mitigating scene corruption from cross-view inconsistencies.
  • Novel functionality: Separating two supervision signals of unequal reliability via a learned per-Gaussian gain field that shares geometry with the LDR base.

In 2D, by contrast, a gain map serves as a learnable storage format or prediction target, and cross-view consistency never arises. Our base-gain representation also enables a novel 3D HDR appearance editing application.

SeHDR already learns a base Gaussian representation and derives an HDR Gaussian representation with shared geometry.

Although both methods share geometry between a base and an HDR representation, they substantially differ in:

  • Base Representation: SeHDR's base stores HDR up to scale (linear radiance at the input exposure). Ours stores LDR appearance, directly supervised by LDR views.
  • HDR Representation: SeHDR derives HDR by fusing exposure-scaled copies of the base via MLPs shared across Gaussians. Ours learns a dedicated per-Gaussian, view-dependent gain, supervised by pseudo-HDR targets.
  • Motivation for Shared geometry: SeHDR shares geometry because its bracketed Gaussians are exposure variants of one scene. We use it to let LDR views anchor the scene while directing pseudo-HDR inconsistencies toward the gain.
  • Functionality: Only our gain field separates HDR enhancement from the LDR base, which enables 3D HDR editing.
AuthorsOpenReview ↗

Response to Reviewer 977U (2/5)

W2 & Q1: Reproduction of SeHDR

In short: The SeHDR paper leaves several implementation details and its HDR normalization undefined. In our reimplementation, we matched its evaluation protocol wherever specified, consistently resolved open choices in its favor, and reported its best results from several rounds of tuning. We will revise the paper to acknowledge the gap to SeHDR's published results and credit SeHDR with its higher numbers. Even taking SeHDR's published results at face value, PriorGS leads on all three metrics (31.22 vs. 25.37 dB PSNR). Details follow.

Thank you for raising this point. SeHDR's code is not public, and its paper reports results only for HDR views of synthetic scenes, so we reimplemented it to compare on all datasets and on LDR re-exposure; all SeHDR numbers in our tables come from this reimplementation. We followed the paper wherever it is specific, including its evaluation protocol (18/17 train/test views; one training run per exposure, averaged over all five; μ-law with μ = 5000). For the details it leaves open, listed below, we consistently chose in SeHDR's favor and, where we tested alternatives, kept the best-performing one.

  • NeEF fusion space: The paper describes NeEF as operating in SH-coefficient space (their Sec. 3, Fig. 2 caption, contributions). Their Eqs. 9-12, however, fuse per-view evaluated colors. Because the fusion is nonlinear, the two readings give different models, and the paper does not specify which one it uses. Our choice: we implemented both and kept the better one: per-view evaluated colors.

  • Bracket exposures: The paper samples additional exposures Δti2,…,Δtin\Delta t_i^2, \dots, \Delta t_i^n (their Sec. 3.2) but does not give nn, their values or range, or the sampling rule: fixed or random, per iteration, or per Gaussian, as the index ii allows. Our choice: each scene's full exposure set. It includes the test exposures which inform the training of NeEF. This favors SeHDR, giving it information no other method receives.

  • CRF selection: SeHDR fixes one DoRF curve per scene, chosen as the best-performing curve on a randomly sampled training image, evaluated by "dynamic range and smoothness" (their Sec. 3.3). The paper does not define these metrics, how they are combined, or how a curve is scored on a single LDR image; the random image choice may also change the selected curve between runs. Our choice: the per-scene CRF learned by full PriorGS. It is the strongest CRF in our ablations (our Table 2, f-h), and PriorGS with it achieves the best re-exposure in our Table 1. Reinhard tone mapping (smooth, with a wide dynamic range, matching SeHDR's two criteria), gamma 2.2 and 2.4, and a log curve all performed worse for SeHDR.

  • HDR normalization and luminance correction: Their Sec. 4.1 computes metrics on μ-law tone-mapped normalized HDR radiance but does not define the normalization: per image or per scene, or whether ground-truth HDR enters it. Our choice: the same per-scene luminance correction as for all methods (our Appendix A.3). Please refer to our response to Q3 of Reviewer EB8P for further details on our per-scene luminance correction mechanism.

  • Training settings: Iterations, learning rates, and resolutions are not reported. Our choice: the configuration shared by all methods (our Appendix A.3).

The reported numbers are the best we obtained over several rounds of tuning.

Gap to the published 25.37 dB. We will revise the paper to acknowledge this gap and credit SeHDR with its higher published results. As normalizations may differ, these results are not strictly comparable to ours; still, PriorGS exceeds them on all three metrics (31.22 / 0.9569 / 0.038 vs. 25.37 / 0.810 / 0.117 PSNR/SSIM/LPIPS).

AuthorsOpenReview ↗

Response to Reviewer 977U (3/5)

W2 & Q2: Analysis for benefit of closed-loop reconstruction across datasets

In short: The closed loop improves results on all datasets, including a new HDR ablation on synthetic scenes, at negligible cost. Its gain depends on how much mismatch between pseudo-HDR targets and input LDR views it corrects: measured directly at the observed exposure, this correction is 2.53 dB on HDR-NeRF-Real but only 1.59 dB on HDR-Plenoxels-Real, explaining the smaller gain at unseen exposures there. Details follow.

Thank you for this question.

New HDR ablation on synthetic scenes. Using HDR GT, we ablate the loop on HDR-NeRF-Syn (all variants in Q3). The loop improves HDR recovery:

MethodPSNR ↑SSIM ↑LPIPS ↓
w/o loop31.010.95200.043
PriorGS31.220.95690.038

Why the gain varies across real datasets. PSNR on test views, without → with the loop:

DatasetObserved exposure (re-exposed)Unseen exposures (Table 2(b))
HDR-NeRF-Real30.08 → 32.61 (+2.53)23.52 → 24.31 (+0.79)
HDR-Plenoxels-Real28.10 → 29.69 (+1.59)22.97 → 23.00 (+0.03)

Without the loop, the HDR follows the pseudo-HDR targets, and the re-exposure loss fits only the CRF. Any mismatch between these targets and the input LDR views that a global CRF cannot absorb thus remains in the HDR. The loop corrects this by refining the HDR, so its effect can be measured directly at the observed exposure. There, it improves PSNR by 2.53 dB on HDR-NeRF-Real but only 1.59 dB on HDR-Plenoxels-Real. The pseudo-HDR targets thus leave less mismatch to correct on HDR-Plenoxels, so its benefit at unseen exposures is smaller.

Necessity. The loop improves results on every dataset at negligible cost and without adding any parameters. Its benefit is also larger than PSNR suggests: On HDR-Plenoxels, LPIPS still improves from 0.099 to 0.092 (Table 2(b)). Qualitatively, removing the loop introduces artifacts and degrades HDR views (Figs. 7, 15).

W2 & Q3: Additional quantitative ablations on HDR-NeRF synthetic data

In short: As requested, we extended our real-scene ablations to synthetic scenes with HDR GT. All ablations trail PriorGS, consistent with our original ablations (Table 2) and qualitative HDR ablations, and further confirming each component's contribution to HDR recovery. Details follow.

We thank the reviewer for this suggestion and will add the following analysis to the paper.

Requested analysis. We agree that HDR GT assesses each component's contribution to HDR recovery most directly. We ran ablations (a)-(e) on HDR-NeRF-Syn against HDR GT, with the same per-scene luminance correction as Table 1:

(ID) DescriptionPSNR ↑SSIM ↑LPIPS ↓
(a) Remove LDR-to-HDR prior29.070.92720.078
(b) Remove reconstruction loop31.010.95200.043
(c) Base-Gain → direct encoding28.520.87350.170
(d) Gain Field → MLP(c, μ, d)28.050.85360.126
(e) Gain Field → MLP(c)30.350.95640.039
PriorGS (Ours)31.220.95690.038

All ablations trail PriorGS on every HDR metric, consistent with the overall trends of our real-scene ablations (Table 2) and with our qualitative ablations (Figures 2, 7, 15-17 and 19). We omit ablation for the CRF variants (f)-(h), as the CRF serves LDR re-exposure, which Table 2 already evaluates directly.

Why we ablated on real scenes. We ablated on real rather than synthetic scenes for their more challenging content and more diverse capture/exposure settings (Sec. 4.3). As they lack HDR GT, we assessed HDR recovery through re-exposure at unseen exposures, which expose the input's saturated (lower exposures) and dark (higher exposures) regions, an established way to evaluate HDR quality (Cao et al. [1]).

[1] Cao et al. "Perceptual assessment and optimization of HDR image rendering." CVPR 2024.

Comments on the paper

2 comments

Thank you for raising your score and for your kind message. We are grateful for your fair questions and comments which have helped us improve the quality of our work and the manuscript. We are happy to offer any further clarifications before the discussion closes.

Best regards,

The Authors

AuthorsOpenReview ↗

Summary of Rebuttal

We sincerely thank all reviewers once again for their constructive comments and insightful questions. We are encouraged that reviewers found our editable gain field novel (EB8P) and meaningful (977U), our performance strong (all), and our motivation (5TVd) and presentation (EB8P, 977U) clear. We have responded to each comment in detail and have finalized the revised manuscript with changes highlighted in color. The revision addresses all reviewer feedback. We summarize the main changes below:

  • Related works: We expanded our related work section to further distinguish our contributions from those of previous works (Sec. 2; Appendix A.1).

  • Same-prior baselines: While PriorGS substantially outperforms existing methods, we also isolated the contribution of our 3D HDR framework by building two baselines that share our HDR reconstruction prior. PriorGS consistently outperforms both, one of which is a 2D HDR baseline built on PriorGS's own source of HDR supervision. PriorGS thus improves on its own HDR supervision rather than merely inheriting it (Sec. 4.1, 4.2, and Figure 20 for qualitative evidence).

  • Extended ablations: We extended our ablations from real scenes to synthetic scenes with HDR GT. The new results are consistent with the original ones and further confirm the contribution of each component of our framework (Sec. 4.3).

  • Further analysis: We added analyses of (1) the robustness of our base-gain representation to cross-view inconsistencies (Appendix A.8); (2) the accuracy of our learned CRFs against GT CRFs (Figure 12; Appendix A.5); (3) region-wise HDR reconstruction (Appendix A.8); and (4) computational efficiency (Appendix A.9).

  • Clarification on SeHDR results: Table 1 now reports SeHDR's published results alongside our reproduction; PriorGS outperforms both on all HDR metrics.

We sincerely appreciate the reviewers' feedback which has helped us improve the quality of our work and manuscript. We look forward to the final evaluation of Reviewers 977U and EB8P based on our responses and the revision, and we are happy to provide any further clarification before the discussion period ends.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.