Pith. sign in

REVIEW 5 major objections 5 minor 47 references

Dual-Output Multi-Exposure HDR Reconstruction via SDR Fusion and Gain Map Inverse Tone Mapping

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read DOME-HDR claims that a single framework can jointly produce a display-ready SDR image and a high-dynamic-range HDR image from three bracketed LDR exposures, reporting the top scores across three multi-exposure HDR benchmarks.

desk verdict Real engineering with a plausible joint SDR+HDR design, but the gain-map math contradicts itself and the training target is undefined; the SOTA claim is not interpretable as written. read the letter →

arxiv 2608.05626 v1 pith:GHR3E33Z submitted 2026-08-06 cs.CV

classification cs.CV
keywords multi-exposureHDRreconstructiongainmapinversetonemappingfusionlatentdiffusionmodelLoRAadaptationdualcross-attentionpriorimagequalityassessment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

DOME-HDR claims that a single framework can solve multi-exposure HDR reconstruction by first generating a clean standard-dynamic-range (SDR) image from three differently exposed photos and then expanding that SDR into a high-dynamic-range (HDR) image with a learned gain map. The point of the two-stage design is that ordinary gain-map methods can only scale what the SDR already contains, so clipped highlights are unrecoverable; here the SDR is synthesized by a diffusion model with exposure-specific attention branches, giving the gain map a more complete base to expand. If the claim holds, HDR imaging no longer needs separate fusion and reconstruction models, and the final SDR is display-ready without choosing a tone-mapping operator. The paper reports state-of-the-art results on the combined Kalantari, Tel, and Challenge123 test sets, with PU-PSNR 44.63 and PU-SSIM 0.9878 in Table 1.

What carries the argument

The load-bearing identity is the gain-map expansion $\hat{I}_{hdr} = (\hat{I}_{sdr}+\epsilon)^{\hat{g}}\hat{Q}_{max}-\epsilon$, where $\hat{g}$ is predicted as a residual over $G_{prior} = \log(\tilde{I}_{hdr}^{prior}+\epsilon)/\log(\hat{I}_{sdr}+\epsilon)$, so that the prior alone already reproduces a saturation-aware blend of the three exposures. HPGM is a U-Net encoder-decoder that consumes an 11-channel input (base SDR, two exposure ratios, two soft saturation masks) and injects the HDR prior at the bottleneck, while Sobel edge maps gate the upsampling. The mechanism's job is to let the synthesized SDR supply detail that a pure multiplicative gain map cannot invent, and to let the network correct a rough radiance prior rather than predict a gain map from scratch.

What would settle it

Retrain AFUNet, the strongest baseline in Table 1, under exactly DOME-HDR's protocol including the merged training splits, merged test splits, and the same $\mu$-law tone mapping for no-reference metrics; if it reaches or exceeds PU-PSNR 44.63 or PU-SSIM 0.9878, the paper's central state-of-the-art claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-exposure HDR reconstruction should be decomposed into two jointly optimized tasks: synthesize a perceptually balanced base SDR from the three LDR exposures, then estimate a per-pixel gain map that lifts the SDR to HDR. The SDR stage builds on the UltraFusion diffusion model, adapted with LoRA, and uses two independent cross-attention branches so under- and over-exposed cues are injected into the mid-exposure anchor without being averaged together. The gain-map stage, HPGM, predicts the gain as a residual over a hand-built HDR prior formed by gamma-linearizing the three exposures, computing exposure ratios, and blending saturation-aware estimates; the final HDR is the gain-expanded SDR scaled by a global factor. With two-phase training, the paper reports the best PU-PSNR, PU-SSIM, LPIPS, DISTS, CLIPIQA, NIQE, and BRISQUE scores on the merged test sets, including PU-PSNR 44.63 versus 44.48 for the previous best baseline.

Load-bearing premise

The state-of-the-art claim rests on the assumption that the published scores of the compared methods were obtained under the same combined training data, test splits, and tone-mapped scoring conditions as DOME-HDR.

Editorial extensions

If this is right

  • A three-bracket burst becomes one inference that returns both a display-ready SDR and an HDR image, so HDR viewing no longer depends on a separate tone-mapping operator.
  • Saturated regions that stump classic gain-map inverse tone mapping can be recovered, because the base SDR is reconstructed by the diffusion stage instead of being taken as fixed input.
  • Separate per-exposure cross-attention branches prevent the averaging of under- and over-exposed cues, which the ablations credit with most of the improvement over the UltraFusion feature-fusion strategy.
  • On the merged Kalantari, Tel, and Challenge123 evaluation, the reported gains are largest on perceptual metrics (LPIPS and DISTS) rather than on PU-PSNR, pointing to improved visual realism alongside fidelity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the authors do not run: replace the diffusion-synthesized SDR with a simpler exposure-fusion SDR and keep HPGM fixed, which would isolate how much of the HDR gain comes from the generative base versus the gain-map network.
  • The full-reference margins over the best baseline are small (0.15 dB PU-PSNR, 0.0003 PU-SSIM), so on many individual scenes the differences may be within run-to-run variation; per-scene error bars would sharpen the comparison.
  • The framework could in principle handle other bracket counts by adding or removing cross-attention branches, but the paper evaluates only the three-exposure setting.
  • Because the HDR prior uses fixed threshold masks, unusual radiometry might benefit from learned masks; the paper does not explore that variant.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes DOME-HDR, a dual-output framework that takes three bracketed LDR images and produces both a display-ready SDR image and an HDR image. Stage 1 uses a LoRA-adapted latent diffusion model, built on UltraFusion, to synthesize the SDR via two separate cross-attention branches conditioned on under- and over-exposed inputs with the mid-exposure as anchor. Stage 2 introduces HPGM, a U-Net that predicts a spatially varying gain map and a global scale, guided by an 'HDR prior' computed as a ratio of logarithms of a normalized linear HDR estimate and the synthesized SDR. The HDR is reconstructed by raising the SDR plus epsilon to the predicted gain-power and subtracting epsilon. The model is trained with an image-level loss (L1 in the mu-law domain plus VGG perceptual loss) and a gain-level loss whose target gain map is not defined. Experiments on merged Kalantari/Tel/Challenge123 test splits report state-of-the-art PU-PSNR, PU-SSIM, LPIPS, DISTS, CLIPIQA, NIQE, and BRISQUE values, with ablations of the cross-attention design, the HDR prior inputs, and the two-phase training strategy.

Significance. If the empirical claims hold, DOME-HDR would be a useful step toward unifying multi-exposure fusion and HDR reconstruction: it produces an SDR output that does not depend on a tone-mapping operator and an HDR output via gain-map expansion, and the dual cross-attention design is a reasonable extension of UltraFusion. The paper also evaluates both full-reference and no-reference metrics and provides several ablations. However, the current manuscript does not provide a reproducible definition of the gain map or its training target, and the SOTA comparison lacks evidence that baseline numbers were obtained under the same merged protocol. These gaps make the significance conditional rather than established.

major comments (5)
  1. [§2.3, Eq. (1)-(2); §3.1 Eq. (3); §3.3 Eq. (9)] §2.3 defines the gain map as the multiplicative ratio f=(H+ε)/(S+ε) and reconstructs HDR as (S+ε)⊙f−ε. In contrast, §3.1 Eq. (3) and §3.3 Eq. (9) reconstruct HDR by exponentiation: (Ŝ+ε)^ĝ−ε and (Ŝ+ε)^{ĝ·Qmax}−ε. These are not algebraically equivalent, and for tone-mapped SDR values in [0,1] they have opposite effects on highlights: a multiplicative factor larger than 1 brightens, while an exponent larger than 1 darkens. The paper therefore uses two incompatible definitions of 'gain map', and it is unclear which one the network is trained to predict. This is an internal inconsistency that blocks reproduction.
  2. [§3.4, Eq. (11)] The gain-level loss uses a 'ground-truth gain map g' and a ground-truth global scale Q_max, but neither is defined anywhere in the manuscript. The three benchmark datasets provide HDR ground truth only; no paired ground-truth SDR is available, so the target (S,g) for the loss is missing unless the authors specify how g is constructed. If g is the ratio from Eq. (1), then Eq. (9) does not invert it; if g is instead the exponent in Eq. (9), Eq. (1) is incorrect. The paper must state the exact formula used to derive g and Q_max from the training data, or the reported results cannot be reproduced.
  3. [§4.1, Table 1] Table 1 compares DOME-HDR with seven previous methods on the merged test splits, but §4.1 does not state whether the baselines were retrained on the combined training set or whether their numbers were taken from the original per-dataset publications. Since the paper deliberately changes the training protocol by combining three datasets, baseline scores from single-dataset papers are not directly comparable. The reported performance margins over AFUNet are small (44.63 vs 44.48 PU-PSNR, 0.9878 vs 0.9875 PU-SSIM), and no error bars, repeated runs, or significance tests are provided. This makes the SOTA claim unsupported as written.
  4. [§3.3, Eq. (8)] The HDR prior G_prior is computed as log(̃I_hdr_prior + ε) / log(Ŝ + ε), where ̃I_hdr_prior is called a 'normalized HDR estimate'. The normalization is never defined, and the ratio-of-logarithms formula is presented without derivation or justification. Because G_prior is both the initialization and the residual base for the final gain map prediction, an undefined normalization and an ad-hoc formula have a direct effect on the reconstructed HDR. The authors should either derive G_prior from the gain-map definition used in training or explicitly present it as an empirically chosen initialization and study its sensitivity.
  5. [§3.4, Eq. (10)] The perceptual loss in Eq. (10) is written as ||φ(I_hdr) − φ(ˆI_hdr)||_1, where φ is VGG19 features. VGG19 expects SDR-range inputs, while I_hdr and ˆI_hdr are linear HDR radiance images that can have values far outside [0,1]. Unless the HDR images are tone-mapped before being passed to φ (which the equation does not indicate), the perceptual loss is ill-posed and may dominate training with unstable gradients. Please specify the exact preprocessing used for φ and, if tone-mapping is applied, write the loss accordingly.
minor comments (5)
  1. [§3.2, Eq. (4)] Eq. (4) uses a temperature τ in the softmax but never defines it; if τ=√C, please state this.
  2. [Fig. 3] The caption repeats 'Tone-mapped Results (Ours)' twice and does not clearly identify which columns are ground truth and which are outputs.
  3. [References] References [32] and [33] are the same paper (Tel et al., 'Alignment-free HDR deghosting with semantics consistent transformer', ICCV 2023); please remove the duplicate.
  4. [§4.5, Table 2] Table 2 lists MEF-LUT and UltraFusion but does not state whether these baselines were retrained under the same three-exposure input protocol; the sentence 'all methods use the same three exposure images as input' needs a concrete citation or a description of the implementation.
  5. [§4.4] The text of §4.4 is vague ('achieves better reconstruction quality compared with previous methods'); please spell out the metric-by-metric margins reported in Table 1.

Circularity Check

1 steps flagged · score 2.0 of 10

No logical circularity: the HDR output is externally supervised against held-out ground truth and is not a rename of fitted inputs; one flagged definitional inconsistency (gain map defined multiplicatively in Eq. 1 but used exponentially in Eqs. 3/9, with ground-truth g in Eq. 11 undefined) blocks reproduction without making the result circular.

  1. self definitional [Sections 2.3-3.4: Eq. (1)-(2) vs. Eqs. (3), (8)-(9), (11)]
    "f(x,y) = (H(x,y) + ϵ)/(S(x,y) + ϵ) (1) ... Îhdr(x,y) = (Îsdr(x,y) + ϵ)^ĝ(x,y) − ϵ (3) ... Gprior = log(Ĩhdr_prior + ϵ)/log(Îsdr + ϵ) (8) ... Lgain = √((ĝ−g)² + ϵ²) + w_qmax||Q̂max − Qmax||₁ (11), where g denotes the ground-truth gain map"

    Eq. (1) defines the gain map as a multiplicative ratio f=(H+ϵ)/(S+ϵ), but the reconstruction actually used in Eqs. (3)/(9) exponentiates: Îhdr=(Îsdr+ϵ)^(ĝ·Qmax)−ϵ. For SDR values in (0,1) the operations act oppositely, so the predicted quantity is not the defined quantity. Gprior in Eq. (8) is the log-ratio which, inserted into Eq. (9) with Δg=0 and Qmax=1, yields Îhdr=Ĩhdr_prior: the zero-residual HDR output collapses to the hand-blended exposure prior of Eq. (7), a function of the inputs alone. The ground-truth gain map g in Eq. (11) is never defined; if g follows Eq. (1), Eq. (9) does not invert it, and if g follows Eq. (8), Eq. (1) is wrong. The chain (SDR→gain map→HDR) is not a well-defined function of the paper's own definitions.

full rationale

Score 2: no significant logical circularity. The central SOTA claim is evaluated against external ground truth (Kalantari, Tel, and Challenge123 test splits, Table 1) with image-level supervision (Eq. 10) against real HDR targets, so the HDR output is neither a rename of fitted inputs nor a byproduct of a self-citation chain. Prior works invoked (UltraFusion [7], GMNet [22], gain-map specifications [1,2,9], benchmarks [14,16,33]) are external to this author team, and no load-bearing self-citation exists. The self-referential construction of Gprior from the model's own SDR output (Eq. 8) and the anchoring of Îhdr to Îsdr (Eqs. 3/9) are design choices of the GM-ITM paradigm, not logical loops, because the training signal (ground-truth HDR, Eq. 10) comes from outside the model. Two flagged defects are correctness or reproducibility risks rather than circularity: (1) the gain map is defined multiplicatively in Eq. (1) but used exponentially in Eqs. (3)/(9), and the ground-truth gain map g in Eq. (11) is undefined, so the claimed gain-map prediction cannot be reproduced or interpreted; (2) Table 1 merges the three test splits but does not state whether baselines were retrained under the identical combined protocol, so the reported margins (44.63 vs. 44.48 PU-PSNR) may reflect evaluation inconsistency. Per hard rule 3, a pipeline benchmarked against external data with independent supervision receives a low circularity score.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's central claim depends on a small set of hand-set hyperparameters (saturation thresholds, mask sharpness, loss weights) and on several modeling assumptions, including the validity of gamma-based linearization, the reliability of the exposure-blending prior, and the expressiveness of the exponentiated-gain representation. The most notable gap is that the ground-truth gain map used in the loss is never defined, so the training target is an implicit assumption.

free parameters (5)
  • Saturation mask thresholds t_hi, t_lo = 0.95, 0.05
    Hand-set thresholds in Eq. (6) that define which regions are treated as highlights and shadows for the HDR prior; they affect the prior and hence the gain map output.
  • Mask sharpness alpha = 50
    Hand-set sigmoid sharpness in Eq. (6); controls the softness of the saturation masks.
  • Loss weights (lambda_perc, w_qmax) = 1e-2, 0.3
    Hand-chosen weights in Eqs. (10) and (11); balance perceptual loss and gain-map/scale supervision.
  • Tone-mapping parameter mu = 5000
    Mu-law compression parameter used in the loss (Eq. 10) and in evaluation; fixed by convention but chosen by the authors.
  • LoRA rank = 8
    Rank of the low-rank adaptation applied to attention projections; a hyperparameter selected by the authors.
assumptions (5)
  • domain assumption Gamma decoding with gamma=2.2 and exposure normalization converts the three LDR inputs into linear radiance estimates (Eq. 5).
    Assumes a simple gamma response and that exposure values ev_k are known and accurate; real camera response functions may deviate.
  • domain assumption The saturation-aware blending in Eq. (7) produces a reliable HDR prior from the three linearized exposures.
    Assumes under-exposed, mid-exposed, and over-exposed images are correctly exposed and aligned such that their weighted sum approximates scene radiance in each region.
  • ad hoc to paper The HDR image can be represented as (SDR + eps)^(g * Q_max) - eps for a spatially varying gain map g and global scale Q_max.
    This exponentiation form is introduced in Eq. (9) without justification; it differs from the multiplicative gain map definition in Eq. (1), and its expressiveness is assumed rather than proven.
  • domain assumption The datasets provide aligned ground-truth HDR images and corresponding SDR references for training.
    Training losses in Eqs. (10) and (11) presuppose access to aligned HDR ground truth and to a gain map g derived from it, but the derivation of g is never specified.
  • domain assumption The frozen U-Net and VAE of the diffusion backbone are suitable for this task after LoRA adaptation of the attention projections.
    The paper assumes the pretrained UltraFusion latent space generalizes to the three-exposure fusion and HDR guidance tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-Output Multi-Exposure HDR Reconstruction via SDR Fusion and Gain Map Inverse Tone Mapping." pith.science (2026). https://pith.science/paper/GHR3E33Z

@misc{pith2026260805626,
  author       = {Pith},
  title        = {Pith review of: Dual-Output Multi-Exposure HDR Reconstruction via SDR Fusion and Gain Map Inverse Tone Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GHR3E33Z}},
  note         = {Machine review of arXiv:2608.05626}
}
read the original abstract

We propose DOME-HDR, a dual-output multi-exposure HDR reconstruction framework that jointly produces a perceptually balanced SDR image and a consistent HDR image via gain map inverse tone mapping. Given three bracketed LDR inputs, DOME-HDR first synthesizes a base SDR using a LoRA-adapted latent diffusion model. A dual cross-attention fusion module injects complementary structural and color cues from the under- and over-exposed images while anchoring on the mid exposure for stability. The synthesized SDR then guides HPGM, our HDR Prior-guided Gain Map network, to predict a spatially varying gain map for reliable dynamic-range expansion. We evaluate on Kalantari, Tel, and Challenge123 using both full-reference and no-reference metrics, where DOME-HDR achieves state-of-the-art HDR reconstruction quality; ablations further confirm the effectiveness of dual cross-attention and SDR-guided gain map estimation.

Figures

Figures reproduced from arXiv: 2608.05626 by the authors.

Figure 1
Figure 1. Comparison of dynamic range expansion strategies. Existing multi-exposure fusion (MEF) [7], multi-exposure HDR reconstruction [18], and gain map-based method [22] operate in disjoint domains and produce partial outputs. In contrast, our framework jointly predicts MEF, gain map, tone-mapped SDR, and HDR using a single framework. For visualization, HDR results are tone-mapped in the PQ domain. HDR reconstruction [5, 6… view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. Given three exposure images, our method first synthesizes a visually balanced base SDR image through a diffusion-based multi￾exposure fusion stage. The fused SDR image is then used to estimate a spatially varying gain map with the proposed HDR prior–guided gain map network. 3 Methodology 3.1 Overview [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison with HDR reconstruction methods. Our method pro￾duces visually pleasing PQ tone-mapped HDR results with improved highlight recovery and richer color appearance compared to existing approaches. 4.4 Comparison on HDR Reconstruction [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with MEF methods. Our method produces more natural fusion with fewer highlight artifacts [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Comparison of different visualization operators (µ-law, PQ, and Reinhard [31]) and our SDR results. Our method directly produces a perceptually balanced SDR image, enabling consistent visualization without additional tone mapping. quality of existing HDR reconstruction…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 46 canonical work pages

  1. [1]

    Adobe: Gain map specification.https://helpx.adobe.com/camera-raw/using/ gain-map.html(2024), accessed: June 27, 2026

  2. [2]

    Apple: Explore hdr rendering with edr.https://developer.apple.com/videos/ play/wwdc2021/10161/(2021), accessed: June 27, 2026

  3. [3]

    In: PCS (2021)

    Azimi, M., et al.: Pu21: A novel perceptually uniform encoding for adapting exist- ing quality metrics for hdr. In: PCS (2021)

  4. [4]

    Canham, T.D., Tedla, S., Murdoch, M.J., Brown, M.S.: Gain-mlp: Improving hdr gain map encoding via a lightweight mlp. In: Int. Conf. Comput. Vis. (2025)

  5. [5]

    In: AAAI (2023)

    Chen, R., Zheng, B., Zhang, H., Chen, Q., Yan, C., Slabaugh, G., Yuan, S.: Hybrid transformer for multi-exposure hdr imaging. In: AAAI (2023)

  6. [6]

    In: IEEE Conf

    Chen, X., Liu, Y., Zhang, Z., Qiao, Y., Dong, C.: Hdrunet: Single image hdr recon- struction with denoising and dequantization. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)

  7. [7]

    In: IEEE Conf

    Chen, Z., Wang, Y., Cai, X., You, Z., Lu, Z., Zhang, F., Guo, S., Xue, T.: Ultrafu- sion: Ultra high dynamic imaging using exposure fusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2025)

  8. [8]

    In: CVPR (2020)

    Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. In: CVPR (2020)

Show all 47 references
  1. [9]

    Google: Ultra hdr image format.https : / / developer . android . com / media / platform/hdr-image-format(2024), accessed: June 27, 2026

  2. [10]

    Hu, E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: Int. Conf. Learn. Represent. (2022)

  3. [11]

    In: IEEE Conf

    Hu, T., Yan, Q., Qi, Y., Zhang, Y.: Generating content for hdr deghosting from frequency view. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)

  4. [12]

    arXiv preprint arXiv:2512.16357 (2025)

    Hu,T.,Zhou,W.,Tu,Y.,Wu,P.,Dong,W.,Yan,Q.,Zhang,Y.:Gmodiff:One-step gain map refinement with diffusion priors for hdr reconstruction. arXiv preprint arXiv:2512.16357 (2025)

  5. [13]

    Jiang, T., Wang, C., Li, X., Li, R., Fan, H., Liu, S.: Meflut: Unsupervised 1d lookup tables for multi-exposure image fusion. In: Int. Conf. Comput. Vis. (2023)

  6. [14]

    ACM Transactions on Graphics (2017)

    Kalantari, N.K., Ramamoorthi, R.: Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics (2017)

  7. [15]

    In: IEEE Conf

    Kim, J., Zhu, Z., Bau, T., Liu, C.: Dcdr-unet: Deformable convolution based detail restoration via u-shape network for single image hdr reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024) 16 J. Kim et al

  8. [16]

    Kong, L., Li, B., Xiong, Y., Zhang, H., Gu, H., Chen, J.: Safnet: Selective alignment fusion network for efficient hdr imaging. In: Eur. Conf. Comput. Vis. pp. 256–273 (2024)

  9. [17]

    IEEE Transactions on Image Processing (2020)

    Li, H., Ma, K., Yong, H., Zhang, L.: Fast multiscale structural patch decomposition for multi-exposure image fusion. IEEE Transactions on Image Processing (2020)

  10. [18]

    Li, X., Ni, Z., Yang, W.: Afunet: Cross-iterative alignment-fusion synergy for hdr reconstruction via deep unfolding paradigm. In: Int. Conf. Comput. Vis. pp. 10666– 10675 (2025)

  11. [19]

    IEEE Transactions on Image Processing (2017)

    Li, Z., Wei, Z., Wen, C., Zheng, J.: Detail-enhanced multi-scale exposure fusion. IEEE Transactions on Image Processing (2017)

  12. [20]

    IEEE Transactions on Image Processing (2014)

    Li, Z., Zheng, J., Zhu, Z., Wu, S.: Selectively detail-enhanced fusion of differ- ently exposed images with moving objects. IEEE Transactions on Image Processing (2014)

  13. [21]

    Liang, P., Jiang, J., Liu, X., Ma, J.: Fusion from decomposition: A self-supervised decomposition approach for image fusion. In: Eur. Conf. Comput. Vis. (2022)

  14. [22]

    Liao, Y., Guan, Y., Xu, R., Li, J., Sun, S., Xiong, Z.: Learning gain map for inverse tone mapping. In: Int. Conf. Learn. Represent. (2025)

  15. [23]

    In: AAAI (2023)

    Liu, R., Li, C., Cao, H., Zheng, Y., Zeng, M., Cheng, X.: Emef: Ensemble multi- exposure image fusion. In: AAAI (2023)

  16. [24]

    In: IEEE Conf

    Liu, Z., Lin, W., Li, X., Rao, Q., Jiang, T., Han, M., Fan, H., Sun, J., Liu, S.: Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)

  17. [25]

    Liu, Z., Wang, Y., Zeng, B., Liu, S.: Ghost-free hdr imaging with context-aware vision transformer. In: Eur. Conf. Comput. Vis. (2022)

  18. [26]

    IEEE Transactions on Image Processing (2017)

    Ma, K., Li, H., Yong, H., Wang, Z., Meng, D., Zhang, L.: Robust multi-exposure image fusion: A structural patch decomposition approach. IEEE Transactions on Image Processing (2017)

  19. [27]

    In: Pacific Graphics (2007)

    Mertens, T., Kautz, J., Van Reeth, F.: Exposure fusion. In: Pacific Graphics (2007)

  20. [28]

    IEEE Transactions on Image Processing (2012)

    Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing (2012)

  21. [29]

    IEEE Signal Processing Letters (2012)

    Mittal, A., Soundararajan, R., Bovik, A.C.: Making a completely blind image qual- ity analyzer. IEEE Signal Processing Letters (2012)

  22. [30]

    Prabhakar, K.R., Srikar, V.S., Babu, R.V.: Deepfuse: A deep unsupervised ap- proach for exposure fusion with extreme exposure image pairs. In: Int. Conf. Com- put. Vis. (2017)

  23. [31]

    In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp

    Reinhard, E., Stark, M., Shirley, P., Ferwerda, J.: Photographic tone reproduction for digital images. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 661–670. ACM (2023)

  24. [33]

    Tel, S., Wu, Z., Zhang, Y., Heyrman, B., Demonceaux, C., Timofte, R., Ginhac, D.: Alignment-free hdr deghosting with semantics consistent transformer. In: Int. Conf. Comput. Vis. (2023)

  25. [34]

    In: IJCAI (2022)

    Wang, H., Ye, M., Zhu, X., Li, S., Zhu, C., Li, X.: Kunet: Imaging knowledge- inspired single hdr image reconstruction. In: IJCAI (2022)

  26. [35]

    In: CVPR (2023)

    Wang, J., Chan, K.C., Loy, C.C.: Clip-iqa: No-reference image quality assessment with clip. In: CVPR (2023)

  27. [36]

    IEEE Transactions on Image Process- ing (2004) DOME-HDR 17

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Process- ing (2004) DOME-HDR 17

  28. [37]

    In: AAAI (2024)

    Wu, G., Fu, H., Liu, J., Ma, L., Fan, X., Liu, R.: Hybrid-supervised dual-search: Leveraging automatic learning for loss-free multi-exposure image fusion. In: AAAI (2024)

  29. [38]

    Wu, S., Xu, J., Tai, Y.W., Tang, C.K.: Deep high dynamic range imaging with large foreground motions. In: Eur. Conf. Comput. Vis. (2018)

  30. [39]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)

    Xu, H., Ma, J., Jiang, J., Guo, X., Ling, H.: U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)

  31. [40]

    IEEE Transactions on Image Processing (2020)

    Xu, H., Ma, J., Zhang, X.P.: Mef-gan: Multi-exposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing (2020)

  32. [41]

    In: IEEE Conf

    Yan, Q., Gong, D., Shi, Q., Hengel, A.v.d., Shen, C., Reid, I., Zhang, Y.: Attention- guided network for ghost-free high dynamic range imaging. In: IEEE Conf. Com- put. Vis. Pattern Recog. (2019)

  33. [42]

    TCSVT (2023)

    Yan, Q., Hu, T., Sun, Y., Tang, H., Zhu, Y., Dong, W., Van Gool, L., Zhang, Y.: Towards high-quality hdr deghosting with conditional diffusion models. TCSVT (2023)

  34. [43]

    TIP (2020)

    Yan, Q., Zhang, L., Liu, Y., Zhu, Y., Sun, J., Shi, Q., Zhang, Y.: Deep hdr imaging via a nonlocal network. TIP (2020)

  35. [44]

    In: IEEE Conf

    Yang, S., Ding, K., Zhang, W., Ma, K.: Maniqa: Multi-dimension attention network for no-reference image quality assessment. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

  36. [45]

    In: ACM Int

    Ye, Q., Xiao, J., Lam, K.M., Okatani, T.: Progressive and selective fusion network for high dynamic range imaging. In: ACM Int. Conf. Multimedia (2021)

  37. [46]

    In: CVPR (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)

  38. [47]

    TIP 21(4), 2318–2323 (2011)

    Zhang, W., Cham, W.K.: Gradient-directed multi-exposure composition. TIP 21(4), 2318–2323 (2011)

  39. [48]

    In: IEEE Conf

    Zhu, P., Sun, Y., Cao, B., Hu, Q.: Task-customized mixture of adapters for general image fusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.