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..
WarpGAN: Warping-Guided 3D GAN Inversion with Style-Based Novel View Inpainting
Authors: Kaitao Huang, Yan Yan, Jing-Hao Xue, Hanzi Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Quantitative and qualitative experiments demonstrate that our method consistently outperforms several state-of-the-art methods. ... 4 Experiments |
| Researcher Affiliation | Academia | 1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, P.R. China 2Department of Statistical Science, University College London, UK |
| Pseudocode | No | The paper describes the methodology in Section 3, detailing the components and their interactions through text and mathematical equations, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Corresponding Author Please visit the Project Page for visualizations and code. ... We submit the code in the supplementary material, and all the datasets used are publicly available. |
| Open Datasets | Yes | Our experiments mainly focus on face datasets. We use the FFHQ dataset [20] and 100K pairs of synthetic data for training. ... To evaluate the generalization ability of our method, we employ the Celeb A-HQ dataset [19] and the multi-view MEAD dataset [40] for testing. ... all the datasets used are publicly available. |
| Dataset Splits | Yes | Datasets. Our experiments mainly focus on face datasets. We use the FFHQ dataset [20] and 100K pairs of synthetic data for training. The synthetic pairs {Isynth s , Isynth t } are generated from EG3D [5], sharing the same latent code wsynth but rendered with different camera poses. To evaluate the generalization ability of our method, we employ the Celeb A-HQ dataset [19] and the multi-view MEAD dataset [40] for testing. We preprocess the images in the datasets and extract their camera poses in the same manner as [5]. ... For the multi-view MEAD dataset, each person includes five face images with increasing yaw angles (front, 30 , and 60 ). We use the front image as input and synthesize the other four views. |
| Hardware Specification | Yes | The inference times (Time) in Table 1 are measured on a single Nvidia Ge Force RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions several software components like EG3D [5], Swin-Transformer [26], Arc Face network [12], Ranger optimizer, and Adam optimizer [22], but does not provide specific version numbers for these or other key software dependencies. |
| Experiment Setup | Yes | For the 3D GAN inversion encoder Ew+, we set the batch size to 4 and train it for 500K iterations on the FFHQ dataset. We use the Ranger optimizer, which combines Rectified Adam [25] with the Lookahead technique [47], with learning rates of 1e-4 for Ew+. The values of λ2, λLPIPS, and λw+ ID in Eq. (3) are set to 1.0, 0.8, and 0.1. For SVINet, we set the batch size to 2 and train it for 300K iterations on both the FFHQ dataset and synthetic data pairs. ... We use the Adam optimizer [22], with learning rates of 1e-3 and 1e-4 for the SVINet and discriminator, respectively. The values of λ1, λP, and λID in Eq. (6) are set to 10.0, 30.0, and 0.1, respectively. The values of λrec, λc, and λadv in Eq. (10) are set to 1.0, 0.1, and 10.0, respectively. |