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

Conference: ECCV 2026
Paper: ECCV Official
Area: Image Generation
Keywords: HDR Reconstruction, Multi-Exposure Fusion, Gain Map, Inverse Tone Mapping, Diffusion Models

TL;DR

DOME-HDR introduces a dual-output multi-exposure HDR reconstruction framework that decomposes the problem into LoRA-adapted latent diffusion SDR fusion and HDR prior-guided gain map estimation (HPGM), concurrently delivering display-ready SDR images and physically accurate HDR radiance within a single pipeline.

Background & Motivation

Real-world natural scenes exhibit luminance variations that span several orders of magnitude between deep shadows and bright specular highlights. Standard dynamic range (SDR) imaging sensors, constrained by physical dynamic range and 8-bit precision, fail to capture this full dynamic spectrum, leading to severe under-exposure or clipped highlights. To alleviate this bottleneck, two distinct avenues of research have developed: multi-exposure fusion (MEF) and high dynamic range (HDR) reconstruction. MEF merges bracketed low dynamic range (LDR) exposures into a single perceptually well-balanced SDR image; recent latent diffusion models have further improved perceptual realism and contrast, yet their outputs remain strictly confined to the LDR domain without recovering scene radiance.

Conversely, multi-exposure HDR reconstruction aims to recover linear radiance maps by aligning and fusing multiple LDR inputs. However, existing HDR reconstruction pipelines suffer from fundamental practical constraints. First, modern generative backbones (e.g., VAEs and diffusion priors) are predominantly pretrained on vast SDR image collections, making their direct adaptation to linear HDR radiance computationally intensive and prone to domain discrepancies. Second, linear HDR representations require post-hoc tone mapping operators (TMOs) for display visualization, rendering final visual fidelity highly variable across display hardware. Gain map-based inverse tone mapping (GM-ITM) offers a backward-compatible alternative by predicting a multiplicative residual between SDR and HDR. However, conventional GM-ITM assumes a single pristine SDR input; when highlights in the SDR image are severely clipped or saturated, a multiplicative gain map cannot synthesize the lost information from nothing. Moreover, naively cascading a standalone SDR generator and a gain map estimator induces severe error propagation from SDR artifacts to gain map predictions.

This paper addresses this fundamental dilemma by bridging the multi-exposure paradigm with the gain map representation. Core idea: disentangle perceptual exposure fusion from HDR luminance expansion by anchoring a dual cross-attention diffusion model on the mid-exposure to synthesize a base SDR, and guiding an edge-gated gain map network (HPGM) with a multi-exposure physical radiance prior to predict residual gain and global scale, jointly producing display-ready SDR and high-fidelity HDR within a unified framework.

Method

Overall Architecture

DOME-HDR takes three bracketed LDR images—under-exposed \(I_{ue}\), mid-exposed \(I_{me}\), and over-exposed \(I_{oe}\)—and maps them to both a base SDR image \(\hat{I}_{sdr}\) and a calibrated HDR image \(\hat{I}_{hdr}\). The framework is decomposed into two interconnected stages: 1. Base SDR Synthesis (Stage 1): Using the mid-exposure \(I_{me}\) as the central structural anchor in the latent space, exposure-specific structural and color cues extracted from \(I_{ue}\) and \(I_{oe}\) are integrated through separate cross-attention branches. The latent representation guides a frozen latent diffusion backbone adapted via low-rank adaptation (LoRA) to synthesize a visually pleasing base SDR image \(\hat{I}_{sdr}\) via the VAE decoder. 2. Gain Map Estimation and Inverse Tone Mapping (Stage 2): A physical HDR prior \(G_{prior}\) is first constructed via gamma linearization and soft saturation masks across the three exposures. The HDR Prior-guided Gain Map network (HPGM) consumes an 11-channel tensor comprising \(\hat{I}_{sdr}\), cross-exposure ratios, and saturation masks. Injected with bottleneck prior features and edge-gated during upsampling, HPGM predicts a residual gain map \(\Delta g\) (yielding \(\hat{g} = G_{prior} + \Delta g\)) and a global scale factor \(\hat{Q}_{max}\), reconstructing the final HDR image via exponential scaling.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
    A["Three-Exposure LDR Inputs: I_ue, I_me, I_oe"] --> B["Stage 1: Dual Cross-Attention SDR Synthesis<br/>Mid-exposure anchor + independent UE/OE guidance"]
    B --> C["Base SDR Image I_sdr"]
    A --> D["Multi-Exposure Physical Radiance Initialization<br/>Gamma linearization + soft saturation masking"]
    D --> E["Physical HDR Prior G_prior"]
    C & E --> F["Stage 2: HPGM Gain Map & Scale Estimation<br/>11-channel input + bottleneck prior injection + edge gating"]
    F --> G["Predicted Gain Map g = G_prior + Δg<br/>and Global Scale Q_max"]
    C & G --> H["Dual Output Delivery<br/>Display-ready SDR + calibrated HDR"]

Key Designs

1. Dual Cross-Attention Feature Injection: Preserving Exposure-Specific Structural Details

In diffusion-based exposure fusion, baseline architectures like UltraFusion merge guidance features from multiple exposures via normalized spatial summation before feeding them into cross-attention. However, when handling three exposures, the under-exposed image preserves sharp high-frequency highlights while the over-exposed image contains low-noise shadow colors; arithmetic averaging dilutes these complementary cues. DOME-HDR introduces two independent cross-attention branches where the latent feature of the mid-exposure \(X_{me}\) serves as the Query, while concatenated structural and color cues from \(I_{ue}\) and \(I_{oe}\) serve as Keys and Values:

\[X_{out} = X_{me} + \text{Conv}_{1\times 1}\left(\text{softmax}\left(\frac{QK^\top}{\tau}\right)V\right)\]

This formulation applies separately to the UE and OE branches, allowing the model to selectively borrow highlight structures and shadow tones without mutual interference. To ensure computational tractability and preserve generative priors, the diffusion U-Net and VAE remain completely frozen; only LoRA adapters with rank 8 are inserted into the attention query and value projection matrices (\(W_q, W_v\)).

2. Physical HDR Radiance Prior Construction: Providing a Stable Inductive Bias

Predicting an unconstrained gain map directly from an SDR image is an ill-posed inverse problem that often produces color shifts and unnatural luminance discontinuities. DOME-HDR exploits the physical properties of the input exposure stack to initialize a linear radiance estimate. Each exposure is first linearized via gamma decoding (\(\gamma = 2.2\)) and normalized by its exposure value \(ev_k\):

\[X_k = \frac{I_k^\gamma}{2^{ev_k}}, \quad k \in \{ue, me, oe\}\]

Using the mid-exposure luminance \(l_{me}\), soft saturation masks for highlights (\(S_{hi}\)) and shadows (\(S_{lo}\)) are calculated using sigmoid functions with thresholds \(t_{hi} = 0.95\), \(t_{lo} = 0.05\), and sharpness \(\alpha = 50\): \(S_{hi} = \sigma(\alpha \cdot (l_{me} - t_{hi}))\) and \(S_{lo} = \sigma(\alpha \cdot (t_{lo} - l_{me}))\). A saturation-aware blended radiance estimate is then formed: \(\hat{I}_{hdr}^{prior} = S_{hi} X_{ue} + (1 - S_{hi} - S_{lo}) X_{me} + S_{lo} X_{oe}\). Finally, the initial gain prior is derived in logarithmic ratio against the base SDR:

\[G_{prior} = \frac{\log(\tilde{I}_{hdr}^{prior} + \epsilon)}{\log(\hat{I}_{sdr} + \epsilon)}\]

By anchoring the gain map to this physically grounded baseline, the subsequent network only needs to learn a subtle residual correction rather than absolute dynamic scaling.

3. HPGM Network and Dual-Output Co-Reconstruction: Edge-Gated Residual Learning

The HPGM network is configured as a U-Net that takes an 11-channel concatenated input: the base SDR \(\hat{I}_{sdr}\), cross-exposure ratios \(R_{ue} = X_{ue}/X_{me}\) and \(R_{oe} = X_{oe}/X_{me}\), and the saturation masks \(S_{hi}, S_{lo}\). At the network bottleneck, a lightweight convolutional encoder extracts prior feature representations \(f_{prior}\) from \(G_{prior}\) and the masks, injecting them into the bottleneck via residual addition. In the decoder, three edge-guided upsampling blocks apply Sobel-filtered edge maps as spatial gates to sharpen transition boundaries and prevent halo artifacts. The network terminates in two heads: a convolutional gain map head predicting the residual \(\Delta g\) (\(\hat{g} = G_{prior} + \Delta g\)), and a global scale head computing the overall HDR dynamic range multiplier \(\hat{Q}_{max}\) via global average pooling and a 2-layer MLP. The final HDR radiance is reconstructed as:

\[\hat{I}_{hdr}(x,y) = \left(\hat{I}_{sdr}(x,y) + \epsilon\right)^{\hat{g}(x,y)} \cdot \hat{Q}_{max}\]

This formulation guarantees that the high-frequency perceptual texture synthesized in the SDR stage is preserved, while dynamic luminance is faithfully expanded.

Loss & Training

The framework is optimized end-to-end with joint image-level and gain-level supervision: 1. Image-Level Loss \(\mathcal{L}_{img}\): Computes the \(L_1\) distance in the \(\mu\)-law compressed domain (\(\mu = 5000\)) alongside a VGG19 perceptual loss (\(\lambda_{perc} = 0.01\)):

\[\mathcal{L}_{img} = \|\mathcal{T}_\mu(I_{hdr}) - \mathcal{T}_\mu(\hat{I}_{hdr})\|_1 + \lambda_{perc} \|\phi(I_{hdr}) - \phi(\hat{I}_{hdr})\|_1\]

1. Gain-Level Loss \(\mathcal{L}_{gain}\): Enforces fidelity on the gain map and global scale: \(\mathcal{L}_{gain} = \sqrt{(\hat{g} - g)^2 + \varepsilon} + w_{qmax} |\hat{Q}_{max} - Q_{max}|\), with \(\varepsilon = 10^{-3}\) and \(w_{qmax} = 0.3\). The total loss is \(\mathcal{L}_{total} = \mathcal{L}_{img} + \mathcal{L}_{gain}\).
2. Two-Phase Training Strategy: In Phase 1, the LoRA-adapted diffusion module and HPGM are jointly trained for 30k iterations on ground-truth SDR pairings. In Phase 2, HPGM is further trained for 30k iterations using the actual MEF outputs synthesized by Phase 1 as the input SDR base. This two-phase approach bridges the domain gap between pristine training targets and generative inference outputs.

Key Experimental Results

Main Results

On the combined test benchmarks of Challenge123, Tel, and Kalantari, DOME-HDR is compared against top-tier multi-exposure HDR reconstruction methods (Table 1 of the paper; NR-IQA metrics evaluated on \(\mu\)-law tone-mapped images):

Method	PU-PSNR (dB) ↑	PU-SSIM ↑	LPIPS ↓	DISTS ↓	CLIPIQA ↑	NIQE ↓	BRISQUE ↓
AHDRNet	41.99	0.9780	0.0438	0.0309	0.3963	3.61	24.78
HDR-Transformer	43.74	0.9806	0.0382	0.0232	0.3982	3.62	24.12
SCTNet	42.92	0.9783	0.0399	0.0246	0.3959	3.56	23.84
SAFNet	42.98	0.9802	0.0344	0.0213	0.4019	3.63	23.12
DiffHDR	42.50	0.9739	0.0464	0.0297	0.4101	3.86	26.15
RFG-HDR	43.16	0.9788	0.0367	0.0239	0.3927	3.59	23.98
AFUNet (Prev. SOTA)	44.48	0.9875	0.0205	0.0162	0.3983	3.68	27.13
DOME-HDR (Ours)	44.63	0.9878	0.0156	0.0147	0.4137	3.56	22.18

In evaluating the standalone perceptual quality of the generated base SDR against specialized multi-exposure fusion (MEF) algorithms (Table 2 of the paper):

Method	MANIQA ↑	CLIPIQA ↑	NIQE ↓	BRISQUE ↓
MEF-LUT	0.4064	0.5147	3.34	36.53
UltraFusion	0.3650	0.4070	3.92	17.01
DOME-HDR (Ours)	0.4129	0.5233	3.37	16.52

Ablation Study

Extensive ablations evaluate each core technical choice:

1. Cascaded Baseline vs. Unified Framework (Table 3) Directly chaining a pretrained UltraFusion model with a pretrained GMNet results in severe performance collapse (PU-PSNR of 22.96 dB and LPIPS of 0.1042). In contrast, DOME-HDR's joint formulation achieves 44.63 dB PU-PSNR and 0.0156 LPIPS, validating the necessity of co-optimizing SDR fusion and gain map prediction.

2. Cross-Attention Design & Prior Input Configurations (Tables 4 & 5)

Experiment / Configuration	PU-PSNR (dB) ↑	PU-SSIM ↑	LPIPS ↓	DISTS ↓	Note
UltraFusion Feature Summation	43.76	0.9876	0.0180	0.0157	Summation causes feature dilution
Dual Cross-Attention (Ours)	44.63	0.9878	0.0156	0.0147	+0.87 dB gain; sharper textures
Prior: SDR only	23.69	0.8487	0.1310	0.0846	No exposure cues; severe failure
Prior: Under + Mid (UE, ME)	25.32	0.8794	0.1126	0.0732	Missing shadow radiance
Prior: Over + Mid (OE, ME)	26.16	0.8920	0.1041	0.0701	Missing highlight recovery
Prior: All Three Exposures	44.63	0.9878	0.0156	0.0147	Complete exposure complement

3. Two-Phase Training Strategy (Table 6) - Phase 1 Only: Yields 40.53 dB PU-PSNR and 0.0162 LPIPS. While visually coherent, gain estimation suffers from the domain gap of generated SDRs. - Phase 2 Only: Achieves 43.26 dB PU-PSNR and 0.0143 DISTS. - Combined Phase 1 + Phase 2: Reaches 44.63 dB PU-PSNR and 0.0156 LPIPS, achieving the best overall quantitative and perceptual scores.

Key Findings

• Physical Exposure Prior Prevents Radiance Collapse: When HPGM is conditioned solely on the base SDR without physical exposure priors, PU-PSNR drops by over 20 dB (23.69 dB vs. 44.63 dB). This confirms that inverse tone mapping from a single SDR fundamentally cannot reconstruct clipped highlight radiance without multi-exposure guidance.
• Elimination of Display Tone Mapping Variability: Across traditional HDR methods (AFUNet, SCTNet, SAFNet), applying different tone mapping operators (\(\mu\)-law, PQ, Reinhard) produces erratic perceptual quality scores (CLIPIQA ranging from 0.35 to 0.39). In contrast, DOME-HDR's Stage 1 directly yields a balanced SDR output achieving 0.4984 CLIPIQA and 19.88 BRISQUE without requiring any external tone mapper.

Highlights & Insights

• Dual Output Decouples Display Constraints: By generating both an SDR image and an HDR gain map, DOME-HDR sidesteps the black-box variability of downstream display tone mappers. Standard SDR displays render the high-quality fusion directly, while HDR displays expand radiance via the gain map, preserving backward compatibility.
• Multi-Exposure Blending as a Strong Inductive Bias: Rather than treating gain map estimation as a purely black-box deep learning task, the authors compute a closed-form radiance estimate via gamma linearization and soft saturation gating, transforming gain prediction into an easier residual learning task (\(\Delta g\)).
• Two-Phase Fine-Tuning Eliminates Generative Domain Shift: Joint end-to-end training often struggles when a downstream network receives imperfect generative outputs during inference. Tuning HPGM on the actual Stage 1 MEF outputs during Phase 2 ensures seamless domain alignment.

Limitations & Future Work

• Diffusion Latency Overhead: Stage 1 utilizes an iterative latent diffusion process. Even with LoRA parameter efficiency, the multi-step denoising process incurs higher inference latency than pure single-forward CNN/Transformer networks, posing challenges for real-time mobile viewfinder integration.
• Sensitivity to Extreme Foreground Motion: Although cross-attention to the mid-exposure mitigates severe ghosting, rapid subject movement across bracketed frames can still introduce local boundary artifacts during the linear prior construction stage.
• Future Directions: Exploring few-step rectified flow or diffusion distillation could dramatically accelerate Stage 1 inference, while incorporating deformable alignment into the prior generation pipeline could enhance deghosting under intense motion.

Related Work & Insights

• vs. UltraFusion (CVPR 2025): UltraFusion casts MEF as guided inpainting but restricts its final output to the 8-bit SDR space and relies on feature summation across exposures. DOME-HDR replaces this with dual cross-attention anchored on the mid-exposure and extends the pipeline to HDR radiance reconstruction via HPGM.
• vs. AFUNet (ICCV 2025): AFUNet uses iterative deep unfolding to synergize alignment and fusion for direct linear HDR prediction. DOME-HDR achieves superior distortion and perceptual fidelity (44.63 dB vs. 44.48 dB PU-PSNR; 0.0156 vs. 0.0205 LPIPS) while providing an explicit, display-ready SDR image.
• vs. GMNet / Gain-MLP (ICLR 2025 / ICCV 2025): Prior GM-ITM approaches operate strictly on a single SDR image, suffering from unrecoverable highlight clipping. DOME-HDR incorporates multi-exposure bracketing into gain map estimation, overcoming the fundamental dynamic range ceiling of single-image ITM.

Rating

• Novelty: ⭐⭐⭐⭐ [Clever unification of diffusion-based multi-exposure fusion and gain map inverse tone mapping with physical radiance priors]
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive evaluations across full-reference PU metrics, no-reference perceptual metrics, operator comparisons, and comprehensive modular ablations]
• Writing Quality: ⭐⭐⭐⭐⭐ [Cohesive narrative structure, clear formulation, and well-designed diagrams and tables]
• Value: ⭐⭐⭐⭐ [Highly practical for real-world HDR photography pipelines, offering backward-compatible SDR and forward-looking HDR delivery]

Related Papers

• [CVPR 2025] LEDiff: Latent Exposure Diffusion for HDR Generation
• [CVPR 2026] FaithFusion: Harmonizing Reconstruction and Generation via Pixel-wise Information Gain
• [ECCV 2026] Correlation-Weighted Multi-Reward Optimization for Compositional Generation
• [ECCV 2026] MapDreamer: Aerial Imagery Conditioned Latent Diffusion For Lane Level Map Generation
• [ECCV 2026] Exposure Bias Can Alleviate Itself via Directional and Frequency Rectification in Flow Matching