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..
BecomingLit: Relightable Gaussian Avatars with Hybrid Neural Shading
Authors: Jonathan Schmidt, Simon Giebenhain, Matthias Niessner
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach in extensive experiments on our dataset, where we consistently outperform existing state-of-the-art methods in relighting and reenactment by a significant margin. We evaluate our method on 4 subjects from our dataset, where our focus lies on relighting and self-reenactment. From the 16 available camera views, we use 15 for training, and hold out the center camera for testing. We further hold out 4 light patterns from training altogether. From the available sequences, we use all scripted sequences for training and use the free sequence for testing. As the test metrics, we use the Peak-Signal-to-Noise Ratio (PSNR), Structural-Similarity-Index-Measure (SSIM) [42] and the Learned Perceptual Image Patch Similarity (LPIPS) [52]. We verify the key components of our method with ablation experiments, which we conduct with the same subjects. A qualitative and quantitative comparison is presented in Figure 6 and Table 3, respectively. |
| Researcher Affiliation | Academia | Jonathan Schmidt Simon Giebenhain Matthias Nießner Technical University of Munich |
| Pseudocode | No | The paper describes the methodology in text and mathematical formulas (e.g., Section 4 'Method' and its subsections), but does not contain any explicit 'Pseudocode' or 'Algorithm' blocks, nor structured code-like procedures. |
| Open Source Code | No | Code and data will be released upon acceptance. Following GDPR, we will set up a download form to manage the distribution. Unfortunately, this is not possible in an anonymous form. |
| Open Datasets | Yes | To address the lack of data, we introduce a new multi-view video dataset of different participants in a light stage setting, which we will make publicly available for research purposes. Overall, our contributions are two-fold: We introduce a novel, publicly available dataset, combining high-resolution, high-framerate, multi-view recordings of different subjects in a calibrated light stage setting. |
| Dataset Splits | Yes | From the 16 available camera views, we use 15 for training, and hold out the center camera for testing. We further hold out 4 light patterns from training altogether. From the available sequences, we use all scripted sequences for training and use the free sequence for testing. |
| Hardware Specification | Yes | We train our avatars at 1100x1604 resolution for 250k iterations with a batch size of 4, which takes approximately 30 hours on a single NVIDIA RTX A6000 GPU. |
| Software Dependencies | No | We implement all networks and optimization logic in Py Torch [27], and write custom GPU kernels for the specular shading using the SLANG.D shading language [3]. Specific version numbers for PyTorch or SLANG.D are not provided. |
| Experiment Setup | Yes | We set {λl1, λSSIM} to {1.0, 0.2} in all experiments. We set {λalpha, λscale, λpos} to {2e 2, 2e 2, 1e 5} in all experiments. We empirically set SH degree m to 6 in all experiments. We train our avatars at 1100x1604 resolution for 250k iterations with a batch size of 4, which takes approximately 30 hours on a single NVIDIA RTX A6000 GPU. |