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 Neural Exposure Fields for View Synthesis
Authors: Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakotosaona, Michael Oechsle, Christina Tsalicoglou, Keisuke Tateno, Jonathan Barron, Federico Tombari
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A thorough evaluation of the proposed system and baselines where we find that our approach improves over the state-of-the-art by more than 55% in MSE. 4 Experiments Data. To evaluate our neural scene representation with the latent exposure conditioning, we report reconstruction results on the HDRNe RF [15] dataset (MIT). To further evaluate our model on real-world and scene-level captures, we also report results on the recently proposed Eyeful Tower dataset v2 [51] (CC BY-NC I 4.0). Finally, we test our model on in-the-wild phone captures from [49] (Apache 2.0) and a large-scale scene from [4] (CC-BY). Metrics. For the exposure reconstruction experiments, we follow prior works [6, 15] and report view synthesis metrics PSNR, SSIM, and LPIPS against the ground truth images for both in-distribution and out-of-distribution exposure values. For evaluating our neural exposure field on room-level scenes, we use the HDR data from the Eyeful Tower [51] dataset that allows us to generate tonemapped images with different exposures. Baselines. For the exposure reconstruction experiments, we compare against state-of-the-art methods Ne RF [26], Zip Ne RF [4], 3DGS [18], Ne RF-W [21], HDRNe RF [15], and HDR-GS [6], where HDR-GS is the only method that further requires the HDR captures as input, while all other methods only use RGB at sampled exposure. For the results on the Eyeful Tower scenes, we report our method and all baselines with the same Zip Ne RF [4] backbone to enable a fair comparison. |
| Researcher Affiliation | Industry | Michael Niemeyer Fabian Manhardt Marie-Julie Rakotosaona Michael Oechsle Christina Tsalicoglou Keisuke Tateno Jonathan T. Barron Federico Tombari Google |
| Pseudocode | No | The paper uses mathematical equations and descriptive text to explain the methodology (e.g., equations 1-7 in Section 3) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code, nor does it provide links to a code repository. The paper refers to previous work's architectures but not its own implementation code. |
| Open Datasets | Yes | Data. To evaluate our neural scene representation with the latent exposure conditioning, we report reconstruction results on the HDRNe RF [15] dataset (MIT). To further evaluate our model on real-world and scene-level captures, we also report results on the recently proposed Eyeful Tower dataset v2 [51] (CC BY-NC I 4.0). Finally, we test our model on in-the-wild phone captures from [49] (Apache 2.0) and a large-scale scene from [4] (CC-BY). |
| Dataset Splits | Yes | For the training set, for each view we sample an exposure value from an exposure set (1/2, 1, 2) so that every pose is only observed once with one exposure. For evaluation, we use the HDR data to generate each test view with all exposures from our exposure set and then produce the target image by applying classical exposure fusion [23] as illustrated in Fig. 6. For the exposure reconstruction experiments, we follow prior works [6, 15] and report view synthesis metrics PSNR, SSIM, and LPIPS against the ground truth images for both in-distribution and out-of-distribution exposure values. |
| Hardware Specification | Yes | We train both models jointly for 10,000 iterations ( approx. 10 min.) on forward-facing and for 25,000 iterations (approx. 30 min.) on room- and apartment-sized scenes on 8 V100 GPUs. |
| Software Dependencies | No | The paper mentions using a modified version of Zip-NeRF [4] as the scene representation but does not specify any software versions (e.g., Python, PyTorch, TensorFlow, CUDA versions) used for implementation. |
| Experiment Setup | Yes | We parameterize our neural exposure field as a fully-connected MLP with ReLU activation and four hidden layers with a hidden dimension of 128. In the Ne RF model, we use a bottleneck vector dimension of 256 per default except for forward-facing scenes where we use a dimension of 15. We train both models jointly for 10,000 iterations ( approx. 10 min.) on forward-facing and for 25,000 iterations (approx. 30 min.) on room- and apartment-sized scenes on 8 V100 GPUs. We set the weighting-related hyperparameters as σexp = 0.05, λexp = 0.1, and λsat = 1. |