Bumblebee: 4D Morphing with Exact Dual-Endpoint Control

We study video-guided 4D morphing. Given a source mesh, a target mesh, and a ref- erence video, the task is to generate a mesh sequence with fixed topology. The first and last frames must exactly match the given meshes, and the intermediate motion must follow the reference video. Existing video-to-4D methods mainly focus on re- constructing and generating dynamic meshes. They do not target exact matching of two given endpoints. Even when keyframes are provided as generation conditions, decoded meshes may still deviate from the given geometry. Directly correcting these deviations vertex by vertex can stretch and twist mechanical parts. We propose Bumblebee, a training-free framework built on a frozen video-to-4D generative model. It requires neither skeletal rigging nor optimization of model parameters for each sequence. The two keyframe meshes serve as observed frames at every denoising step to guide the generation of intermediate motion. We use positional dif- ferences between decoded meshes and the given keyframes as geometric feedback. This feedback corrects the trajectory frame by frame, ensuring exact matching at both endpoints. For mechanical assets, we correct motion through whole-part translation and rotation, then compensate for the remaining endpoint deviations. This satisfies the endpoint requirements while reducing stretching and twisting of the parts. Experiments on Motion80 and ActionBench show that Bumblebee outper- forms the compared video-driven generation and shape interpolation baselines in geometry and appearance. Experiments on mechanical assets further show that part rigidity constraints better preserve part shapes than direct vertex-wise correction. §