This work presents a novel framework for video-to-4D generation that creates high-quality dynamic 3D content from single video inputs. It introduces a Direct 4DMesh-to-GS Variation Field VAE to encode canonical Gaussian Splats (GS) and their temporal variations into a compact latent space. Building on this, a Gaussian Variation Field diffusion model is trained with a temporal-aware Diffusion Transformer, conditioned on input videos and canonical GS. The model demonstrates superior generation quality and remarkable generalization to in-the-wild video inputs.
We present a novel framework for video-to-4D generation that creates high-quality dynamic 3D content from single video inputs. Direct 4D diffusion modeling is extremely challenging due to costly data construction and the high-dimensional nature of jointly representing 3D shape, appearance, and motion. We address these challenges by introducing a Direct 4DMesh-to-GS Variation Field VAE that directly encodes canonical Gaussian Splats (GS) and their temporal variations from 3D animation data without per-instance fitting, and compresses high-dimensional animations into a compact latent space. Building upon this efficient representation, we train a Gaussian Variation Field diffusion model with temporal-aware Diffusion Transformer conditioned on input videos and canonical GS. Trained on carefully-curated animatable 3D objects from the Objaverse dataset, our model demonstrates superior generation quality compared to existing methods. It also exhibits remarkable generalization to in-the-wild video inputs despite being trained exclusively on synthetic data, paving the way for generating high-quality animated 3D content.
For detailed installation instructions and how to run a minimal inference example, please refer to the
GitHub repository.
1# Clone the repository
2git clone https://github.com/ForeverFancy/GVFDiffusion.git
3cd GVFDiffusion
4
5# Setup environment and dependencies
6. ./setup.sh --new-env --basic --xformers --flash-attn --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrast
7
8# Run a minimal inference example
9accelerate launch --num_processes 1 inference_dpm_latent.py --batch_size 1 --exp_name /path/to/your/output --config configs/diffusion.yml --start_idx 0 --end_idx 2 --txt_file ./assets/in_the_wild.txt --use_fp16 --num_samples 2 --adaptive --data_dir ./assets/ --num_timesteps 32 --download_assets --in_the_wild
1@article{zhang2025gaussian,
2 title={Gaussian Variation Field Diffusion for High-fidelity Video-to-4D Synthesis},
3 author={Zhang, Bowen and Xu, Sicheng and Wang, Chuxin and Yang, Jiaolong and Zhao, Feng and Chen, Dong and Guo, Baining},
4 journal={arXiv preprint arXiv:2507.23785},
5 year={2025}
6}
