Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning Differential Pyramid Representation for Tone Mapping
Authors: Qirui Yang, Yinbo Li, Yihao Liu, Peng-Tao Jiang, Fangpu Zhang, cheng qihua, Huanjing Yue, Jingyu Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on four diverse datasets: HDR+ Burst Photography [43], HDRI Haven, HDR Survey [44], and the UVTM video dataset [21]. The HDR+ dataset, commonly used for HDR reconstruction and tone mapping research, includes 675 image sets for training and 248 for testing at both 480p and 4K resolutions, following the preprocessing method of Zeng et al. [2]. For HDRI Haven, we collected 570 HDR image pairs from the website HDRI Haven and created corresponding ground truths using several software and toning tools combined with selection algorithms. These images cover a wide range of scenes, such as indoor and outdoor, including natural light/artificial light, multi-contrast, sunrise/sunset, urban/nature, daytime/nighttime, and so on. We select 456 image sets for training and 114 for testing. The HDR Survey dataset consists of 105 real HDR images, with no ground truth, and is one of the benchmarks for HDR tone mapping evaluations [21; 22]. Lastly, the UVTM video dataset, also with no ground truth, includes 20 real captured HDR videos. Note that the HDR Survey and UVTM video datasets are for real-world testing only. |
| Researcher Affiliation | Collaboration | Qirui Yang 1 Yinbo Li 1 Yihao Liu 2 Peng-Tao Jiang 3 Fangpu Zhang 1 Qihua Cheng Huanjing Yue 1 Jingyu Yang 1 1 School of Electrical and Information Engineering, Tianjin University, China 2 Shanghai Artificial Intelligence Laboratory 3 vivo Mobile Communication Co., Ltd |
| Pseudocode | No | The paper describes the methods in Section 3 and provides a framework overview in Figure 3, but it does not include any explicitly labeled pseudocode or algorithm blocks. The procedural steps are explained in narrative text. |
| Open Source Code | No | After acceptance of the paper, we will open up the data and code. |
| Open Datasets | Yes | We evaluate our method on four diverse datasets: HDR+ Burst Photography [43], HDRI Haven, HDR Survey [44], and the UVTM video dataset [21]. The HDR+ dataset... For HDRI Haven, we collected 570 HDR image pairs from the website HDRI Haven... The HDR Survey dataset consists of 105 real HDR images... Lastly, the UVTM video dataset, also with no ground truth, includes 20 real captured HDR videos. |
| Dataset Splits | Yes | The HDR+ dataset, commonly used for HDR reconstruction and tone mapping research, includes 675 image sets for training and 248 for testing at both 480p and 4K resolutions, following the preprocessing method of Zeng et al. [2]. For HDRI Haven, we collected 570 HDR image pairs from the website HDRI Haven and created corresponding ground truths using several software and toning tools combined with selection algorithms. These images cover a wide range of scenes, such as indoor and outdoor, including natural light/artificial light, multi-contrast, sunrise/sunset, urban/nature, daytime/nighttime, and so on. We select 456 image sets for training and 114 for testing. |
| Hardware Specification | Yes | We implement the model using Py Torch on an RTX 3090 GPU and optimize using the Adam W optimizer [48], with β1 = 0.9, β2 = 0.99, and an initial learning rate of 1 10 4. |
| Software Dependencies | No | The paper mentions 'Py Torch' and 'Adam W optimizer [48]' but does not provide specific version numbers for these software components. It also refers to 'VGG19 network [49]' but not as an ancillary software dependency with a version. |
| Experiment Setup | Yes | We implement the model using Py Torch on an RTX 3090 GPU and optimize using the Adam W optimizer [48], with β1 = 0.9, β2 = 0.99, and an initial learning rate of 1 10 4. |