GraphAvatar: Compact Head Avatars with GNN-Generated 3D Gaussians

Authors: Xiaobao Wei, Peng Chen, Ming Lu, Hui Chen, Feng Tian

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments to demonstrate the advantages of Graph Avatar, surpassing existing methods in visual fidelity and storage consumption. The ablation study sheds light on the trade-offs between rendering quality and model size. We evaluate our model using two challenging datasets: Ne RFBlend Shape (abbreviated as NBS data) and INSTA. Tab. 1 displays the metric comparisons against baseline methods, with separate sections for the INSTA and NBS datasets.
Researcher Affiliation Collaboration 1Institute of Software, Chinese Academy of Sciences 2University of Chinese Academy of Sciences 3Intel Labs China EMAIL
Pseudocode No The paper describes the methodology using prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/ucwxb/Graph Avatar
Open Datasets Yes We evaluate our model using two challenging datasets: Ne RFBlend Shape (abbreviated as NBS data) and INSTA. The NBS dataset (Gao et al. 2022) comprises monocular videos from eight subjects... the INSTA dataset (Zielonka, Bolkart, and Thies 2022a) includes data from ten subjects...
Dataset Splits Yes The NBS dataset (Gao et al. 2022) comprises monocular videos from eight subjects, with the last 500 frames of each subject s video designated as the test set. Similarly, the INSTA dataset (Zielonka, Bolkart, and Thies 2022a) includes data from ten subjects, with the final 350 frames of each sequence reserved for testing.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments or training the models.
Software Dependencies No The paper does not explicitly mention any specific software dependencies or library versions (e.g., Python, PyTorch, CUDA versions) used for the experiments.
Experiment Setup Yes We initiate with a warm-up stage for parameter initialization. We treat the target actor as a static scene and utilize vanilla 3D Gaussian Splatting (3DGS) to produce pseudo Gaussians... This warm-up phase is rapid, consisting of 10,000 iterations. ... We supervise the rendered image If using a combination of L1 loss, SSIM (Structural Similarity Index) loss, and LPIPS (Learned Perceptual Image Patch Similarity) loss... with λ = 0.2, λLP IP S = 0.1. ... Our final loss function is: L = λf Lf + λc Lc + λw Lw (12) where λf = 1.0, λc = 0.1 and λw = 0.1. Graph Avatar is trained with the final loss until convergence.