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..
High Resolution UDF Meshing via Iterative Networks
Authors: Federico Stella, Nicolas Talabot, Hieu Le, Pascal Fua
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on diverse 3D models demonstrate that our method produces significantly more accurate and complete meshes than existing approaches, particularly for complex geometries, enabling UDF surface extraction at higher resolutions where traditional methods fail. |
| Researcher Affiliation | Academia | Federico Stella1, Nicolas Talabot1, Hieu Le2, Pascal Fua1 1CVLab, EPFL 2UNC Charlotte 1{firstname.lastname}@epfl.ch, EMAIL |
| Pseudocode | No | The paper describes the network architecture and iterative refinement process using mathematical formulas (Eq. 1, 2, 3) and descriptive text, but it does not include a distinct block explicitly labeled 'Pseudocode' or 'Algorithm' with structured steps. |
| Open Source Code | No | The code used in the paper will be made publicly available in case of acceptance. |
| Open Datasets | Yes | We test our method on five different 3D categories, including garments from MGN [1]; cars, chairs, and planes from Shape Net [3]; and natural scenes from the 3DScene dataset [47], using four different neural UDF architectures. |
| Dataset Splits | Yes | For garments, cars, chairs, and planes, we train a traditional auto-decoder following [27, 36]. To showcase the capabilities of our method on downstream tasks and more precise single-shape neural architectures, we perform surface reconstruction from point clouds by training CAP-L [46]1 on 3D scenes and cars, and Diff UDF [11] on cars. In the supplementary material, we provide additional results using a softplus-based auto-decoder and meshing results obtained from ground-truth UDFs. We use 300 garments and the first 20 samples each for cars, chairs, and planes. As the MGN garments are simpler and exhibit less variety than the other shapes, we use a lower resolution for them. For 3D scenes, we use three scenes from the official CAP-UDF [46] repository2. (...) For each mesh, 20k points are uniformly sampled within the volume, while 400k points are sampled near the surface. To obtain the surface points, 200k points are first uniformly distributed on the surface, and then small amounts of Gaussian noise are added. Gaussian noise with a mean of 0 and a standard deviation of 0.005 is applied to the first 200k surface points, and noise with a mean of 0 and a standard deviation of 0.0005 is added to the remaining 200k points. |
| Hardware Specification | Yes | The times were measured on an NVIDIA A100 GPU. (...) The training takes around 2 hours on an NVIDIA A100-40G GPU. |
| Software Dependencies | No | The network is trained using the Adam optimizer [16] with a learning rate of 5e-4 for 50 epochs. |
| Experiment Setup | Yes | Our network architecture consists of 2 fully connected hidden layers, with 1024 nodes each, and an output layer with 128 outputs. The input layer accepts UDF values and gradients at the 8 cell corners, making up 32 inputs. Additionally, 128 inputs per cell are needed to enable multiple iterations. We consider the current cell and the 6 cells that share a face with it, for a total of 7 * 128 additional inputs, which brings the total number of input nodes to 928 and the total number of trainable weights to around 2.1M. Each layer, except for the final one, is followed by a leaky ReLU activation function with a negative slope of 0.01. The output layer is followed by a sigmoid activation before being used as input for the next iteration, and by a softmax function for the cross entropy loss. The network is trained using the Adam optimizer [16] with a learning rate of 5e-4 for 50 epochs. (...) For our auto-decoder experiments we used the traditional auto-decoder architecture proposed in Deep SDF [27]. The auto-decoder network consists of 12 layers, each with 1024 hidden nodes and ReLU activations, and latent codes of size 512. It is trained using L1 loss, without regularization or Fourier encoding, for 10k epochs with a batch size of 16. To focus the network’s capacity on the surface, the UDF is clamped to 0.1. The Adam optimizer [16] is used with learning rates of 0.0005 for the model and 0.001 for the latent codes, with learning rate decay applied at epochs 1600 and 3500 by a factor of 0.35. |