Diffusion4D: Fast Spatial-temporal Consistent 4D generation via Video Diffusion Models

Authors: HANWEN LIANG, Yuyang Yin, Dejia Xu, hanxue liang, Zhangyang "Atlas" Wang, Konstantinos N Plataniotis, Yao Zhao, Yunchao Wei

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method surpasses prior state-of-the-art techniques in terms of generation efficiency and 4D geometry consistency across various prompt modalities.
Researcher Affiliation Academia Hanwen Liang1 , Yuyang Yin2 , Dejia Xu3, Hanxue Liang4, Zhangyang Wang3, Konstantinos N. Plataniotis1, Yao Zhao2, Yunchao Wei2 1University of Toronto, 2Beijing Jiaotong University, 3University of Texas at Austin, 4University of Cambridge
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We will release the code and the dynamic 3D assets idx of the dataset for reproduction of the results.
Open Datasets Yes We curate a large-scale, high-quality dynamic 3D dataset sourced from the vast 3D data corpus of Objaverse-1.0 [10] and Objaverse-XL [9].
Dataset Splits No The paper mentions a 'test set' but does not specify a 'validation set' or its split.
Hardware Specification Yes We use a valid batch size of 128 and train on 8 NVIDIA A100 GPUs.
Software Dependencies No The paper mentions specific pre-trained models like Video MV [58], Model Scope T2V [42], and I2VGen-XL [56], but does not provide specific version numbers for general software dependencies (e.g., Python, PyTorch).
Experiment Setup Yes We train our 4D-aware video diffusion model for 6k iterations with a constant learning rate of 3 10 5. We use a valid batch size of 128 and train on 8 NVIDIA A100 GPUs. During the sampling stage, we use DDIM [37] sampling with sampling step 50, and w1 = 7.0 and w2 = 0.5 in classifier-free guidance.