GaussianFlow connects the dynamics of 3D Gaussians to resulting pixel velocities. This Gaussian flow is obtained by splatting 3D Gaussian dynamics into the image space, allowing for efficient and end-to-end differentiable generation of dense Gaussian flow. The process enables direct dynamic supervision from optical flow, which significantly benefits 4D dynamic content generation and 4D novel view synthesis with Gaussian Splatting. The method resolves the color drifting issue commonly observed in 4D generation and improves the visual quality of extensive experiments, demonstrating its effectiveness.

The implementation details involve the use of optical flow to supervise the Gaussian flow, allowing for the efficient tracking of motions between consecutive frames. The method is implemented in CUDA, with a focus on balancing speed and effectiveness by using a specific number of Gaussians along each pixel ray. Datasets used for evaluation include the Consistent4D Dataset and the Plenoptic Video Dataset, which contain both synthetic and real-world monocular videos for testing the proposed method.