UltraFusion HDR: AI-Generated HDR Algorithm Captures Detail and Balances Exposure

High‑dynamic‑range (HDR) imaging traditionally relies on exposure fusion of multiple shots, but it works only when the exposure difference is modest (3‑4 EV). Larger gaps cause motion artifacts, ghosting, and loss of detail, especially in dynamic scenes such as moving subjects or back‑lit conditions.

Problem Statement

Photographers often discard images that are severely under‑exposed, over‑exposed, or suffer from motion‑induced ghosting. Existing HDR pipelines cannot reliably merge pairs with exposure differences greater than 4 EV, limiting their usefulness in real‑world shooting scenarios.

Proposed Solution: UltraFusion HDR

A joint team from Shanghai AI Lab, The Chinese University of Hong Kong, and Zhejiang University introduced the "书生·浦像超高动态成像算法" (UltraFusion HDR). The method augments classic exposure fusion with a generative AI model (AIGC) to achieve ultra‑high dynamic range imaging.

Key Technical Steps

Collect two images captured from the same viewpoint but with markedly different exposures.

Use a generative image model to provide prior knowledge about scene structure and color distribution.

Apply a multi‑stage processing pipeline that first aligns the images, then performs robust fusion that tolerates up to 9 EV exposure gaps.

Integrate an adaptive tone‑mapping module that learns mapping functions from the generative model, preserving natural color and contrast.

Employ a guided image restoration framework to maintain fidelity and avoid texture artifacts introduced by the generative model.

How It Overcomes Traditional Limitations

By leveraging the generative model’s prior, UltraFusion HDR can resolve large exposure mismatches and suppress motion‑induced ghosting. The algorithm automatically compensates for alignment errors and lighting variations, producing a single HDR image that retains fine details, reduces noise, and eliminates artifacts even when the exposure difference reaches 9 EV.

Results and Demonstrations

Experiments show that, with a 9 EV exposure gap, the fused output remains natural‑looking, with vivid colors and preserved texture. In dynamic photography scenarios—where subjects move or lighting changes—the method generates consistent HDR results without the typical ghosting seen in conventional pipelines.

Sample images illustrate the before‑and‑after effect: two raw shots with extreme under‑ and over‑exposure are merged into a single high‑quality photograph that retains detail in shadows and highlights.

Practical Access

Home page: https://ultrafusion.openxlab.org.cn/home

Interactive demo: https://openxlab.org.cn/apps/detail/OpenImagingLab/UltraFusion

HuggingFace space: https://huggingface.co/spaces/iimmortall/UltraFusion

The approach demonstrates how generative AI can extend the capabilities of existing imaging hardware without requiring new sensors, offering a cost‑effective way to boost camera performance for both professional and consumer users.