HDRI Poster

SIGGRAPH 2026 Posters

Promoting LDR Panoramas to HDRI

with Rendering-Space Losses

We recover clipped illumination in casual panoramas by training a video diffusion model with losses that measure what matters: the light an HDRI produces.

Zhengming Yu1,3 Li Ma1 Mingming He1 Jingdong Zhang3 Jionghao Wang3 Ning Yu1,2 Paul Debevec1,2

1 Eyeline Labs  ·  2 Netflix  ·  3 Texas A&M University

Watch the demo Abstract View poster
Overview

Optimize the light,
not only the pixels.

HDRIs are the standard light sources for image-based lighting, yet casual LDR panoramas clip precisely where the most important illumination lives. Saturated highlights erase light intensity, color, and direction—errors that become obvious when the panorama lights a virtual scene.

Our method promotes a clipped panorama to linear HDRIs. A fine-tuned video diffusion model reconstructs saturated regions, while diffuse spherical-harmonic and glossy GGX losses directly supervise its image-based lighting behavior.

Approach

Rendering-space supervision

Two complementary lighting objectives guide the model from a clipped panorama to a useful HDR light source.

01 Clipped LDR Panorama + saturation mask
02 HDRI-tuned diffusion model Single-frame masked inpainting
03 Clean light source Linear HDR panorama

Low-frequency response

Diffuse SH loss

Compares 27 second-order spherical-harmonic coefficients in log space, directly supervising irradiance and global energy.

Ldiffuse = Δ(log(1 + cSH))²

Higher-frequency response

Glossy GGX loss

Compares prefiltered specular lobes over 64 × 128 HDRI maps, preserving sharp highlights and reflection response.

Lglossy = ‖Δ(log(1 + EGGX))‖²
Results

Light that survives rendering.

The recovered panoramas match ground-truth lighting across exposure, dominant direction, diffuse shading, and specular reflection.

01 / Exposure recovery

Highlights return as linear HDR light source —not a flat white patch.

Across a six-stop exposure sweep, the output restores structured light sources that remain visible and correctly placed at darker exposures.

Exposure sweep comparing clipped input panoramas with recovered outputs
Input, output, and ground truth HDRIs with color, intensity ratio, and direction error maps
02 / Accuracy

Low color and direction error

The reconstructed dominant light closely matches the ground truth in intensity, chromaticity, and angular position.

Relighting comparison between clipped input, our recovered HDRI, and ground truth
03 / Image-based lighting

Faithful diffuse and glossy response

Relit objects recover the illumination strength, shadowing, and reflections missing from the clipped input.

Demo

See the recovered light in video.

Input, prediction, and ground truth are compared across panorama exposure sweeps and rendered scenes.

Citation

Promoting LDR Panoramas to HDRI with Rendering-Space Losses

SIGGRAPH 2026 Posters

@inproceedings{yu2026promoting,
  title={Promoting LDR Panoramas to HDRI with Rendering-Space Losses},
  author={Yu, Zhengming and Ma, Li and He, Mingming and Zhang, Jingdong and Wang, Jionghao and Yu, Ning and Debevec, Paul},
  booktitle={Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Posters},
  pages={1--3},
  year={2026}
}