Ultra-Fast Inverse Tone Mapping via Gain Map-based LUT

Yuanshen Guan, Ruikang Xu, Chang Chen, Yinuo Liao, Dehua Song, Fenglong Song, Zhiwei Xiong

Keywords: Inverse Tone mapping, Gain Map, Lookup Table, HDR

TL;DR: GMLUT enables real-time 4K ITM in 6.2 ms with +1.4 dB gain, supported by new datasets and a gainmap-based LUT framework.

Abstract:

We aim to introduce Look-Up Tables (LUTs), a highly efficient approach, for ultra-fast inverse tone mapping (ITM).
However, as LUT size scales exponentially with increasing bit-depth, it remains challenging to employ dense sampling for high bit-depth accuracy.
This inevitably introduces quantization artifacts and degrades the fidelity of ITM.
To address this issue, we propose GMLUT, which encodes high-bit-depth HDR information into a low-bit-depth learnable Gain Map (GM), thereby facilitating the application of LUTs.
Nevertheless, since the LUT alone can only perform global mapping, it is insufficient to address local tone-mapping degradations in practical scenarios.
Thus, we devise three closely coordinated operators to address the limitation of LUTs: (a) bilateral grids for local adaptation, (b) image-adaptive LUTs for SDR-to-GM translation, and (c) a lightweight neural modulator for GM refinement.
In addition, we construct a synthetic dataset of over 8,000 4K SDR–GM pairs together with a real-capture test set to support the training and evaluation of GMLUT.
Experiments demonstrate that GMLUT outperforms prior state-of-the-art lightweight ITM methods by +1.4 dB in PQ-PSNR while reducing inference time by 70\%. Remarkably, it processes 4K inputs in only 6.2 ms, achieving significant gains in both accuracy and efficiency.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 9436