No Pose?  No Problem

• Given an unposed source video, See4D spline-interpolates a trajectory of virtual camera poses and estimate per-frame depth to lift each frame into 3D.
• Depth-guided forward warping produces intermediate latents which, together with source latents, are processed by our view-conditional diffusion model.
• The model is further augmented with spatiotemporal attention and noise-adaptive conditioning in an auto-regressive framework, enabling the synthesis of the desired target-view sequence.