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

Junyeong Ahn^1*, Geonwoo Kim^2*, Kinam Kim^1*, Kihwi Kim³, Hoonjin Jung³, Hyojin Jang¹, Jaegul Choo¹

¹KAIST AI ²Inha University ³FLIPTION

Paper arXiv Code Dataset

TL;DR: Given a few reference images of an object, OmniShowcase generates a video of that same object along diverse camera paths, even full 360° orbits, keeping its identity and 3D shape consistent by conditioning the video model on a turntable of surface-normal maps reconstructed from the references.

Orbit 360°

A snowman in a wizard hat and black cloak, hugging a small bird, standing on a floating ice block among glaciers and snow-capped mountains … + [Orbit 360°]

Abstract

Generating videos of a specific subject from a few reference images remains challenging when the camera reveals surfaces that are not visible in the input. Early subject-to-video (S2V) methods condition on a single reference image and must therefore infer the subject's unseen appearance and geometry as the viewpoint changes. Recent methods extend S2V to multiple reference views, providing more information about the subject from different viewpoints, but they do not explicitly specify the 3D structure that connects these observations. In addition, existing training data do not guarantee that viewpoint changes preserve the same underlying 3D subject. We present OmniShowcase, built on two components. First, a fully automated data synthesis pipeline produces training videos along diverse camera trajectories, with viewpoint variation driven by underlying subject geometry, even spanning full 360-degree orbits. Second, a geometry-conditioning method provides the video model with a turntable of surface-normal maps reconstructed from the reference views, without requiring temporal alignment with the generated frames. We construct Showcase-4K, a corpus of 3D-asset-based videos covering twelve camera motions across 88 trajectory configurations with reconstruction-derived surface-normal conditioning, and ShowcaseBench, a held-out evaluation set that ships the source meshes and camera parameters used to render each video. Experiments show state-of-the-art subject identity and 3D consistency while maintaining competitive video quality and motion.

Method

Four reference views are reconstructed in 3D and rendered as a surface-normal turntable. References and normals are concatenated with the target span and read in context by a LoRA-tuned image-to-video diffusion transformer.

Automated 3D-consistent data synthesis pipeline — **Data synthesis pipeline.** Blender renders the target orbit and four reference views from one curated 3D asset. The target is composited into a prompt-driven scene, and the four views are reconstructed and rendered as the normal turntable used as the condition.

Architecture overview — **Architecture.** The reference images and the normal condition enter the reference pathway of the diffusion transformer, with no frame-wise alignment between the normal turntable and the generated video.

Beyond the turntable

The normal condition is a fixed yaw sweep, but the generated camera is not bound to it. The target follows whichever trajectory the prompt names.

Arc-Up

The wooden teddy bear statue on a majolica tile ledge of an Amalfi Coast terrace, among lemon trees, with pastel cliffside houses and a turquoise sea below … + [Arc-Up]

Comparison with baselines

Same references, same prompt and same trajectory for every method.

ShowcaseBench

Arc-Down

Phantom

MAGREF

MV-S2V

3DreamBooth

Ours

The wooden teddy bear statue on the porch of a snow-covered log cabin at dusk, with string lights, firewood, a lantern and snow-laden pines … + [Arc-Down]

NAVI (real-world objects)