LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh
Authors: Jing Wen, Alex Schwing, Shenlong Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed approach on the challenging THuman2.0, XHuman and AIST++ data. Our approach reconstructs an animatable representation from sparse inputs in less than 1s, renders views with 95.1FPS at 1024 1024, and achieves PSNR/LPIPS*/FID of 24.65/110.82/51.27 on THuman2.0, outperforming the state-of-the-art in rendering quality. |
| Researcher Affiliation | Academia | Jing Wen, Alexander G. Schwing & Shenlong Wang University of Illinois Urbana-Champaign EMAIL |
| Pseudocode | No | The paper includes figures (e.g., Figure 7 and 8) illustrating modules and their connections, and describes procedures using mathematical equations and descriptive text. However, it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block, nor does it present any structured steps formatted like code. |
| Open Source Code | No | The paper states: "Please visit the project webpage1 for more qualitative results;" and provides a URL: "https://wenj.github.io/LIFe-Go M/". This is a project webpage for qualitative results, not an explicit statement of code release or a direct link to a code repository. |
| Open Datasets | Yes | We validate our approach on THuman2.0 (Yu et al., 2021), XHuman (Shen et al., 2023) and AIST++ (Li et al., 2021) quantitatively. |
| Dataset Splits | Yes | We follow the experimental setup of GHG (Kwon et al., 2024) and split the dataset into 426 subjects for training and 100 subjects for evaluation. We adopt the training and evaluation protocol of Actors Ne RF (Mu et al., 2023). Specifically, we use subjects 1-15 and 21-30 for training and leave out subjects 16-20 for evaluation. |
| Hardware Specification | Yes | As mentioned, reconstruction needs less than one second and rendering runs at 95.1 FPS on one NVIDIA A100 GPU. |
| Software Dependencies | No | The paper mentions: "We use Adam as the optimizer." and "We use Res Net-18 (He et al., 2016) with Image Net pretrained weights as the image encoder." but does not provide specific version numbers for these or other software components. |
| Experiment Setup | Yes | We set λper = 1.0, λM = 5.0 and λlap = 100 in Eq. (12) on THuman2.0 and λper = 1.0, λM = 0 and λlap = 100 in Eq. (12) on AIST++. We use the SSIM loss in THuman2.0 and the LPIPS loss in AIST++ following the baselines. We use Adam as the optimizer. On THuman2.0, the learning rates of the image encoder and the rest of the model are 1e 4 and 5e 5 respectively. On AIST++, we set the learning rate of all parameters to 5e 5. We optimize the model for 200K iterations on THuman2.0 and 100K iterations on AIST++. |