OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images

OmniShowcase. Omnidirectional Subject-Consistent Video Generation from Multi-View Images. By Junyeong Ahn and six others, from KAIST AI, Inha University, and one other institution. Published on the twenty-sixth of October.

Picture a short video of one particular object, made from a few photos of it. The hard moment comes when the camera swings round and shows a surface none of the photos captured. Early subject-to-video methods, called S2V, condition on a single reference image, so as the viewpoint moves they have to infer the subject's unseen appearance and geometry. More recent methods take several reference views, which gives the model more information about the subject from different angles. Two things are still missing. Those methods do not explicitly specify the 3D structure that connects the observations to one another. And existing training data do not guarantee that a change of viewpoint preserves the same underlying 3D subject.

OmniShowcase answers both gaps, one component for each. The data gap is met by a fully automated synthesis pipeline that produces training videos along diverse camera trajectories, where the viewpoint variation is driven by the subject's underlying geometry. The structure gap is met by a geometry-conditioning method that hands the video model an explicit 3D signal, rather than leaving it to guess shape from appearance.

On the data side, the pipeline's trajectories can range widely, even spanning full three hundred and sixty degree orbits around the subject. On the conditioning side, the signal is a turntable of surface-normal maps, reconstructed from the reference views. A surface-normal map records which way each part of the surface faces, so a turntable of them shows the subject's shape from all around. Importantly, this turntable does not need to be temporally aligned with the frames being generated. The model receives the geometry as a reference, not as a frame-by-frame script.

Two named resources come out of this. Showcase-4K is a corpus of videos built from 3D assets, covering twelve camera motions across eighty-eight trajectory configurations, each with surface-normal conditioning derived from reconstruction. ShowcaseBench is a held-out evaluation set, and it ships with the source meshes and camera parameters used to render each video, so outputs can be checked against the true geometry.

Experiments report state-of-the-art subject identity and 3D consistency. That comes while video quality and motion remain competitive, so the gains in keeping the subject recognisable and geometrically coherent are not bought by giving up how the video looks or moves.

The thread running through the paper is that consistency from every angle has to be taught and specified, not hoped for. Training data where the viewpoint changes follow real geometry teaches it. A reconstructed turntable of surface normals specifies it. Together they let the camera go all the way round a subject that stays the same thing throughout.