CVPR 2026 Submission · Supplementary Project Page
ProjFlow is a training-free projection sampler for flow-matching motion priors. It casts spatial motion control tasks as linear inverse problems and enforces linear spatial constraints exactly while preserving natural motion. This page collects qualitative examples and comparisons to state-of-the-art methods.
ProjFlow satisfies the constraints of any spatial motion-control task that can be formulated as a linear inverse problem.
A kinematics-aware metric spreads corrections coherently across the body during flow-based sampling.
Control is applied to a pretrained text-to-motion prior without retraining or inner-loop optimization.
All four applications shown below can be formulated as linear inverse problems: trajectory following, 2D-to-3D lifting, relative position constraints, and looped motion, all solved by ProjFlow within a single zero-shot framework.
We compare spatial accuracy for single-joint and multi-joint trajectory control. Each card fixes one joint or a set of joints to a target trajectory and shows OmniControl, MaskControl, and ProjFlow under the same control signal.
Text prompt: "a person runs forward in an S path"
Text prompt: "A person jumps and kicks a football in the air with their head"
Text prompt: "the person is boxing with their left hand and throws multiple punches."
Text prompt: "a person puts hands on the armrest"
Text prompt: "a person stands with both feet on the ground, kicks once with his left foot."
Text prompt: "a person side steps left and right."
Text prompt: "a person crosses their arms for chest fly"
Text prompt: "a person snakes forward while the right hand waves up and down"
For the 2D-to-3D reconstruction task, the input is initial 2D keypoints and their trajectories under a known orthographic camera. We compare Sketch2Anim with ProjFlow, which enforces the projection constraints exactly while preserving natural 3D motion.