PaperlayerView exact revision on arXiv

Source-linked AI summary

ExpandDiff: Dynamic Range Expanding Diffusion for Single-Image HDR Reconstruction

Mehmet Emre andıran, Zhuoqian Yang, Liying Lu, Mathieu Salzmann, Sabine Süsstrunk

arXiv:2609.39624v1cs.CVeess.IV

TL;DR

Single-image HDR reconstruction must infer detail lost to variable shadow and highlight clipping while preserving visible LDR content. ExpandDiff uses dynamic clipping synthesis and spatially guided diffusion to reconstruct both clipping directions, improving SI-HDR results over the strongest evaluated competitor.

  • Problem

    Single-image HDR reconstruction must infer clipped shadow and highlight detail while preserving visible content, although the missing detail cannot be uniquely determined from the LDR image.

    Source: §1 · p. 1

  • Method

    ExpandDiff randomly samples shadow and highlight clipping levels for training guidance, then uses spatially guided pixel-space diffusion to predict perceptually encoded HDR.

    Source: §1 · p. 1, §3.1 · p. 2

  • Results

    ExpandDiff-B exceeds the strongest evaluated competing method by a paired mean of 3.43 dB in gain-aligned PU21-PSNR, and ExpandDiff-P exceeds it by 7.34 dB under two-sided clipping.

    Source: §1 · p. 1

  • Takeaways & Limitations

    Across SI-HDR and additional clipping conditions, broadly sampled training transfers beyond its nominal degradation, while benchmark-oriented training can further improve matched performance.

    Source: §5 · p. 4

    ExpandDiff cannot know a scene’s absolute brightness, and its reconstructions weaken over large contiguous clipped regions.

    Source: §5 · p. 4

Abstract

from arXiv · show

Single-image HDR reconstruction requires inferring missing detail while preserving the visible content of an LDR image. Differences in sensor dynamic range and exposure cause LDR images to lose varying amounts of information in shadows and highlights. We present ExpandDiff, a conditional diffusion pipeline that jointly reconstructs clipped shadows and highlights. To account for this variation, we introduce Dynamic Clipping Synthesis (DCS), which randomly samples shadow and highlight clipping percentiles when constructing training inputs from HDR targets. A pixel-space diffusion model guided by spatially-adaptive normalization then predicts perceptually encoded HDR through a bounded output head, reconstructing both clipping directions in one sampling trajectory. On the SI-HDR benchmark, ExpandDiff variants improve HDR reconstruction accuracy by 3.43 dB in PU21-PSNR over the strongest evaluated competing method, and by 7.34 dB under two-sided clipping. The code and supplementary material are available at https://memreandiran.github.io/expanddiff/.

1. INTRODUCTION

Single-image HDR reconstruction must infer detail lost to variable shadow and highlight clipping while preserving visible content, an ill-posed task because the missing detail is not uniquely determined by the LDR image. ExpandDiff combines variable-clipping training with spatially guided, perceptually encoded diffusion and improves benchmark reconstruction results.

1. INTRODUCTION

  • LDR images can lose different amounts of shadow or highlight information as sensor dynamic range and exposure vary, making the missing detail impossible to determine uniquely.Earlier approaches emphasize saturated highlights or reconstruct HDR through intermediate exposure brackets.Source: §1 · p. 1
  • Dynamic Clipping Synthesis randomly samples shadow and highlight clipping levels when generating LDR guidance from HDR targets, exposing one model to varied information loss.The real input’s clipping amount is unknown at inference.Source: §1 · p. 1
  • Spatially-adaptive normalization injects LDR features into the diffusion network, conditioning pixel-space prediction rather than a conditioning or compressed latent space.The adapted RGB-guided RAW-Diffusion backbone predicts HDR RGB images.Source: §1 · p. 1
  • PU21-encoded HDR targets allocate more bounded output range to dark intensities, keeping the loss from being dominated by highlights.The model uses Gaussian diffusion and reconstructs shadows and highlights jointly in one 24-step sampling trajectory.Source: §1 · p. 1
  • ExpandDiff-B exceeds the strongest evaluated competing method by a paired mean of 3.43 dB in gain-aligned PU21-PSNR; ExpandDiff-P exceeds it by 7.34 dB under two-sided clipping.The paper presents the integrated pipeline and its evaluation across clipping distributions as the contribution.Source: §1 · p. 1

2. RELATED WORK

Prior single-image HDR methods recover exposure loss through direct reconstruction or synthesized exposure brackets, while diffusion approaches condition on latent representations or combine separate denoisers. ExpandDiff instead predicts one HDR image in pixel space and jointly reconstructs both clipping directions in one sampling trajectory.

2. RELATED WORK

  • Earlier single-image HDR approaches reconstruct saturated highlights, use multiscale features, feature masking, or camera-pipeline reversal, while other methods synthesize and merge exposure brackets.These approaches use synthetic degradation and recovery of under- and overexposure as training objectives.Source: §2 · p. 1
  • Unlike diffusion methods using encoded input latents or separate highlight and shadow denoisers, ExpandDiff predicts one HDR image in pixel space for both clipping directions.ExpoCM [16] also jointly handles different exposure regimes, using exposure-aware consistency trajectories for one-step inference.Source: §2 · p. 1

3. METHOD

ExpandDiff trains conditional diffusion to recover a PU21-encoded HDR target from synthesized LDR guidance, with the same trajectory reconstructing both clipping directions. Its method varies clipping during training, uses a bounded perceptual prediction domain, and injects spatial image guidance throughout the network.

3. METHOD

  • The method constructs a PU21-encoded HDR target, synthesizes LDR guidance with Dynamic Clipping Synthesis, and trains conditional diffusion to recover the target; inference decodes one guided sampling trajectory to linear HDR.The guidance remains an LDR input while the diffusion model predicts in the encoded HDR domain.Source: §3 · p. 2

3.1. Dynamic Clipping Synthesis

  • DCS samples shadow and highlight clipping percentiles qlo from 0–10% and qhi from 0–30%, with scalar thresholds shared across channels and no thresholds or masks supplied to the network.The percentile ranges span near-unclipped to heavily clipped inputs across scenes with different intensity distributions.Source: §3.1 · p. 2

3.2. PU21 Prediction and the Bounded Head

  • PU21 supervision makes pixelwise errors count according to visibility, and the tanh head bounds the encoded prediction before inverse encoding returns HDR.PU21 gives dark intensities more of the output range and improves the head’s sensitivity to shadow errors.Source: §3.2 · p. 2

3.3. Spatially Conditioned Diffusion Backbone

  • Spatially-adaptive normalization injects multiresolution LDR features into the diffusion U-Net, keeping image structure available throughout reconstruction without a VAE latent HDR prediction.The model uses 24 DDIM steps, and the network reconstructs both clipping directions in the same trajectory.Source: §3.3 · p. 2
  • The diffusion network predicts the clean normalized PU21 target from a Gaussian-noised target and LDR guidance, using equally weighted absolute and squared errors.The logarithmic error term used by RAW-Diffusion is omitted because the target is already perceptually encoded.Source: §3.3 · p. 2

4. EXPERIMENTS

The experiments evaluate ExpandDiff on SI-HDR highlight-only inputs and inputs with simultaneous shadow and highlight clipping, then test transfer and component ablations. The benchmark-oriented variant leads on C95, while the DCS-trained variant leads on Cp.

4.1. Evaluation Setup

  • Evaluation covers 181 SI-HDR scenes: C95 uses benchmark highlight-only inputs, while Cp imposes shadow and highlight clipping at the midpoints of P’s training percentile ranges.C95 has 8.17% of pixels with a channel at the ceiling and none at zero; every Cp image contains both clipping types.Source: §4.1 · p. 3

4.2. Metrics and Alignment

  • Metrics include reference-fidelity measures, no-reference PIQE, and crop-statistic FID-R; scores use unclipped-pixel gain alignment, with separate ground-truth-fitted CRF correction.The CRF fit applies tone-and-color correction independently to each method.Source: §4.2 · p. 3

4.3. SI-HDR Benchmark

  • On C95, ExpandDiff-B reaches 22.61 dB gain-aligned PU21-PSNR, beating ExpandNet by 3.43 dB; on Cp, ExpandDiff-P reaches 31.21 dB gain-aligned PSNR, beating MaskHDR by 7.34 dB.Across the two variants, the paper reports the best values across all eight columns.Source: Table 1 · p. 3, §4.3 · p. 3, §4.4 · p. 3

4.4. Clipping Severity and Distribution Transfer

  • At clipping percentiles (10, 30), P retains a 4.76 dB PSNR advantage and 0.33 JOD HDR-VDP-3 advantage; on LEDiff’s degradation, B leads the strongest competitor by 2.44 dB PSNR and 0.30 JOD.These tests use upper-end clipping for P and an independently defined degradation for LEDiff.Source: §4.4 · p. 4
  • The P/B ranking reverses across input distributions: B leads P on seven metrics at C95, while P leads B on seven at Cp; at 75k steps, P reaches 21.38 dB PSNR on C95, below B’s 22.61 dB.The reported checkpoints otherwise differ in training duration and learning rate.Source: §4.4 · p. 4

4.5. Target Encoding and Output Head

  • On Cp, removing PU21 reduces gain-aligned PSNR by 2.00 dB, removing tanh by 1.01 dB, and removing both by 4.37 dB.Each component improves PSNR with either setting of the other; the ablation does not separate perceptual error weighting from output sensitivity.Source: Table 2 · p. 4, §4.5 · p. 4
  • The ablation’s gains are metric-dependent: the unbounded PU21 model has slightly lower FID-R, while the unbounded linear model has the lowest PIQE.The full model leads the CRF-corrected fidelity metrics.Source: §4.5 · p. 4
  • Inference takes 3.19 s per 512 × 512 image on an NVIDIA A100 80 GB with 24 DDIM steps; 48 or 100 steps show no clear PSNR improvement.Other image sizes are supported after reflection padding to multiples of 64, subject to available memory.Source: §4.5 · p. 4

5. CONCLUSION

ExpandDiff combines Dynamic Clipping Synthesis, spatial LDR guidance, and perceptually encoded diffusion for joint shadow and highlight reconstruction. Results across SI-HDR and additional clipping conditions show transfer beyond nominal degradation, while large contiguous clipped regions remain challenging.

5. CONCLUSION

  • Across SI-HDR and additional clipping conditions, broadly sampled training transfers beyond its nominal degradation, while a benchmark-oriented distribution improves matched performance.The target ablation also supports PU21 with a bounded head for reconstruction fidelity.Source: §5 · p. 4
  • ExpandDiff couples Dynamic Clipping Synthesis, spatial LDR guidance, and perceptually encoded diffusion to reconstruct shadows and highlights jointly.Source: §5 · p. 4
  • Like other methods, ExpandDiff cannot determine a scene’s absolute brightness.Source: §5 · p. 4
  • Reconstruction weakens over large contiguous clipped regions.Source: §5 · p. 4
Loading 2609.39624v1…