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..
PBR-SR: Mesh PBR Texture Super Resolution from 2D Image Priors
Authors: Yujin Chen, Yinyu Nie, Benjamin Ummenhofer, Reiner Birkl, Michael Paulitsch, Matthias Niessner
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate that our method achieves state-of-the-art performance in both texture fidelity and rendering quality, facilitating applications such as relighting, which are beyond the capabilities of traditional image SR methods. |
| Researcher Affiliation | Collaboration | 1 Technical University of Munich 2 Intel Labs |
| Pseudocode | No | The paper describes the methodology in text and provides a pipeline overview figure (Figure 2), but does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | We will release the assets under a reasonable license. (Neur IPS Paper Checklist, Question 13) |
| Open Datasets | Yes | Datasets. Since there is no established benchmark on mesh PBR texture super resolution, we collect a set of PBR meshes with rich texture information [1, 2, 4] (Artist-designed Dataset)... Additionally, we evaluate on 32 AI-generated PBR-textured meshes sourced from the Hyper3D commercial platform [3] (AI-generated Dataset). [1] Cgtrader models. https://www.cgtrader.com. [2] The gltf v2.0 sample models. https://github.com/Khronos Group/gl TF-Sample-Assets. [3] Hyper3d models. https://hyper3d.ai/. [4] Poly haven models. https://polyhaven.com/models. |
| Dataset Splits | Yes | For evaluation, we treat the downloaded high-resolution textures as ground truth. If paired low-resolution textures are provided, we use them directly as input to our method. Otherwise, we generate low-resolution inputs by applying a Gaussian blur followed by bicubic downsampling on the high-resolution textures. ... For training data generation, all texture maps are downsampled to 1024 x 1024 using bicubic interpolation. From these, we randomly extract 480 x 480 patches as high-resolution targets and their corresponding 120 x 120 downsampled versions as low-resolution inputs. This process yields a total of 24,000 LR-HR texture pairs, which are used to fine-tune the 4x PBR super-resolution model. |
| Hardware Specification | Yes | On a single NVIDIA A6000 RTX GPU, optimizing a mesh with PBR textures takes around 30 minutes with 2K-to-8K resolution and less than 8 minutes with 1K-to-2K resolution offline on average. |
| Software Dependencies | No | Our implementation leverages the differentiable renderer from [19] to produce rendered images from selected viewpoints. These images are then super-resolved using the same image SR model to create pseudo-GTs for optimization. Unless explicitly specified, the same environment lighting setup is used for both optimization and evaluation. The optimization uses the Adam optimizer with a constant learning rate of 1e-4. |
| Experiment Setup | Yes | The mesh is normalized to a unit size during the optimization process. The camera is positioned at a distance of 3.25 units with a field of view (FoV) of 10 degrees. We use 750 views in total, sampled from 15 different elevations, with 50 views evenly distributed at each elevation. ... The optimization uses the Adam optimizer with a constant learning rate of 1e-4. In each iteration, we use a batch size of 4, which corresponds to 4 viewpoints. We stop the iterative optimization after 2000 iterations. ... during optimization, the λpix is set to 100 and λreg is 0.5 in the robust pixel-wise loss term. The weighting map is optimized in resolution of 64x64 ... We assign different weights to the channels of the PBR texture maps in the PBR consistency loss function: we set the weight to 1.0 for the diffuse w_Kd, roughness w_Kr, and normal map w_Kn, and 0.1 for the metallic map w_Km. The SSIM part is weighted by λssim = 10 for all. The overall PBR consistency loss is given a weight λpbr = 10 and the PBR TV loss is weighted by λtv = 0.5. |