Generalizable One-shot 3D Neural Head Avatar

Authors: Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, Jan Kautz

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that the proposed approach generalizes well to unseen validation datasets, surpassing SOTA baseline methods by a large margin on head avatar reconstruction and animation.
Researcher Affiliation Industry Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, Jan Kautz https://research.nvidia.com/labs/lpr/one-shot-avatar
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes https://research.nvidia.com/labs/lpr/one-shot-avatar
Open Datasets Yes Training datasets. We train our model using a single-view image dataset (FFHQ [26]) and two video datasets (Celeb V-HQ [74] and RAVDESS [37]).
Dataset Splits No The paper mentions using training datasets (FFHQ, Celeb V-HQ, RAVDESS) and evaluation datasets (Celeb A, HDTF testing split), but does not specify the exact train/validation/test splits (e.g., percentages or sample counts) for all datasets used in their own training process to allow full reproduction of the data partitioning.
Hardware Specification Yes We implement the proposed method using the Py Torch framework [45] and train it with 8 32GB V100 GPUs.
Software Dependencies No The paper mentions software like "Py Torch framework [45]", "Seg Former [59]", and "GFPGAN [58; 51]" but does not specify their version numbers, which are necessary for reproducible dependency descriptions.
Experiment Setup Yes The first training stage takes 6 days, consisting of 750000 iterations. The second training stage takes 2 days with 75000 iterations. Both stages use a batch size of 8 with the Adam optimizer [30] and a learning rate of 0.0001.