UV-free Texture Generation with Denoising and Geodesic Heat Diffusion
Authors: Simone Foti, Stefanos Zafeiriou, Tolga Birdal
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments on two datasets, the chairs of Shape Net [9] and the Amazon Berkeley Objects (ABO) dataset [15]. As we can observe from the quantitative results reported in Tab. 1, our method significantly outperforms Point-UV Diffusion across all metrics. Ablation Studies. We here perform multiple ablations to examine how much each model component, conditioning, and choice contributes to the overall performance. |
| Researcher Affiliation | Academia | Simone Foti Stefanos Zafeiriou Tolga Birdal Department of Computing Imperial College London |
| Pseudocode | No | The paper describes its methods in detail through text and diagrams (e.g., Figure 3, Figure 4, Figure 9), but it does not include a formal pseudocode block or algorithm listing. |
| Open Source Code | Yes | Our code and pre-trained models are available at github.com/simofoti/UV3-Te D. |
| Open Datasets | Yes | We conduct experiments on two datasets, the chairs of Shape Net [9] and the Amazon Berkeley Objects (ABO) dataset [15]. |
| Dataset Splits | Yes | In both cases, we operate a 90 : 5 : 5 split between train, test, and validation sets. |
| Hardware Specification | Yes | We run our models on a single Nvidia A100 with 40GB of dedicated memory. |
| Software Dependencies | No | The paper states, 'We implement our method using Py Torch [37], Pytorch Geometric [21] and Diffusers [52]', but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We use 32 sihks, K = 128 eigenvalues, and a mixed-LBO weighting of ϱ = 0.05. We train our models using the Adam W [30] optimiser for 400 epochs on chairs and 250 on ABO, with a learning rate of 1e 4 and a cosine annealing with 500 warmup iteration steps. We use T = 1, 000 DDPM timesteps, S = 250 farthest point samples in the attention layers, and P = 5, 000 target PDS samples. Our batch size is set to 8 on Shape Net chairs and to 6 on ABO. |