Gcd

GCD's core contribution is teaching Stable Video Diffusion to accept relative camera pose, which turns a video generator into something closer to a controllable virtual camera. The results page shows this pays off in three places. First, amodal completion: the model inpaints occluded scene parts, with the team calling out a paper towel roll and a brown bucket reappearing when the viewpoint is raised. Second, object permanence: blue and red ducks vanish behind a hand and teabox, and a falling brown shoe is hidden by a purple bag, yet the output continues their dynamics, shape and appearance accurately. Third, driving: from a single RGB sensor the model synthesizes a top-down-and-forward view that keeps an out-of-frame white car and two pedestrians visible, and can even output per-pixel semantic segmentation from the novel viewpoint. Two variants are offered — Gradual interpolates the camera path and the team reports it performs better overall, including at large camera movements, while Jumpy displaces the camera directly and hallucinates more, especially for moving objects. The honesty here matters: the driving clips come from the synthetic ParallelDomain engine and real-world demos use the unseen TRI-DDAD dataset, and the authors state zero-shot generalization is highly challenging and not their focus. It is trained on synthetic multi-view RGB-D video data only. For a research group this is a well-documented, openly released asset with code, models and datasets. For anyone wanting a shippable product, expect to invest in adaptation, a GPU budget and deep learning expertise, and accept that outputs are hallucinated rather than geometrically exact.