UltraFusion HDR: How AIGC Enhances Dynamic Imaging to Capture Detail and Balance Exposure

Overview

AIGC can not only create content from nothing but also be stacked with HDR technology to rescue poorly exposed photos, delivering results that closely resemble human‑eye perception. Shanghai AI Lab, together with researchers from the Chinese University of Hong Kong and Zhejiang University, introduced the UltraFusion HDR algorithm (also called 书生·浦像超高动态成像算法) that merges exposure‑fusion techniques with a generative large‑model to achieve ultra‑high dynamic‑range imaging.

Limitations of Conventional HDR

Traditional HDR relies on exposure fusion of multiple shots, but it works reliably only when exposure differences are modest (3–4 EV). In real‑world scenes the exposure gap often exceeds 4 EV, and moving subjects introduce motion artifacts and lighting‑angle inconsistencies, causing severe detail loss.

UltraFusion HDR Architecture

The proposed solution adds a generative image model as a prior to the fusion pipeline. The multi‑stage process first aligns two images captured from the same viewpoint with different exposures, then uses the large‑model‑driven information‑fusion module to handle large exposure gaps and dynamic scenes. Adaptive tone‑mapping learns a natural colour mapping, while a guided‑image‑restoration framework preserves fidelity and prevents the texture‑shifting often introduced by generative models.

Performance Highlights

Even when the two input photos differ by up to 9 EV, UltraFusion HDR can accurately merge them, producing a natural‑looking, detail‑rich image that eliminates noise and motion‑induced artifacts. For moving subjects, varying lighting angles, or partial occlusions, the algorithm still generates a consistent high‑dynamic‑range result, effectively turning “waste shots” into publishable photographs.

Another test shows that with two photos taken under different subject positions and lighting, UltraFusion HDR produces a single motion‑photography image that retains realistic detail and a consistent subject.

Practical Implications

Because the algorithm improves imaging performance without changing hardware, photographers—both professionals and amateurs—can upgrade image quality simply by uploading two differently exposed shots to the UltraFusion HDR service. The system is accessible via the project home page, an online demo, and a HuggingFace Space.

Home page: https://ultrafusion.openxlab.org.cn/home<br/>Demo: https://openxlab.org.cn/apps/detail/OpenImagingLab/UltraFusion<br/>HuggingFace: https://huggingface.co/spaces/iimmortall/UltraFusion