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

OmniShowcase generates omnidirectional subject-consistent videos from multi-view reference images by combining a 3D-asset-based data synthesis pipeline with surface-normal turntable conditioning reconstructed from the references.

Subject-to-video generation from a single reference image must infer surfaces and geometry that are not visible in the input, so generated videos can lose appearance and 3D consistency as the camera moves. Recent multi-view extensions provide more reference information but do not explicitly specify the 3D structure connecting those views. Existing training data also do not guarantee that viewpoint changes preserve the same underlying 3D subject.

OmniShowcase reconstructs the subject in 3D from a small set of reference views and renders the reconstruction as a turntable of surface-normal maps. It conditions a pretrained video diffusion model on both the reference images and this normal sequence, without requiring temporal alignment between the normal and target sequences. The training data pipeline renders target orbits and four reference views from curated 3D assets with Blender, composites the target into a prompt-driven scene by video inpainting, and reconstructs the four views to render the surface-normal turntable. The model injects the reference images and normal condition into the reference pathway of a LoRA-tuned image-to-video diffusion transformer. Training uses Wan2.1 I2V-14B-480P with LoRA rank 128, 25 normal-condition frames and 80 target frames at 480×832 resolution.

• On ShowcaseBench, OmniShowcase reports DINO 0.799, CLIP 0.925, MEt3R 0.190 and GT-CD 0.0824, compared with 3DreamBooth DINO 0.707 and GT-CD 0.1717 and MV-S2V DINO 0.688. §4.2
• On NAVI, OmniShowcase reports DINO 0.828, CLIP 0.918, MEt3R 0.249 and GT-CD 0.0975, compared with MV-S2V DINO 0.770 and GT-CD 0.1816. §4.2
• Removing geometry conditioning changes DINO from 0.799 to 0.752 and GT-CD from 0.0824 to 0.2092 on ShowcaseBench. §4.3
• In the user study with 44 participants and 24 videos, OmniShowcase is preferred on every criterion, taking between 70.9% and 79.4% of the votes while no baseline passes 12.7%. §B
• When training clips are scored as generations against paired Uni3C clips, the 3D-asset render reports DINO 0.8030, MEt3R 0.1871 and GT-CD 0.1323 versus Uni3C DINO 0.7167, MEt3R 0.3052 and GT-CD 0.1409. §E
• Showcase-4K contains 4,000 videos covering 2,593 assets in 620 categories. §G.1

• The conclusion states that the background sometimes looks synthetic.
• The conclusion states that subjects that are not rigid across the four references are not handled.
• Appendix D shows the model assumes four reference views; fewer views are tested, and the full four-view setting performs best.

• Single-view subject-to-video generation leaves surfaces not visible in the input unconstrained, which motivates multiple reference views and explicit geometry. §1
• The data synthesis pipeline renders target orbits and four reference views from 3D assets, then composites the target into a prompt-driven scene by video inpainting. §3.2
• Showcase-4K covers twelve camera motions across 88 trajectory configurations. §Abstract
• Conditioning uses a turntable of surface-normal maps reconstructed from the reference views without temporal alignment to generated frames. §3.3
• Reference images and the normal condition are injected into the reference pathway of a LoRA-tuned image-to-video diffusion transformer. §3.3
• Each training sample contains a 25-frame normal-condition segment and an 80-frame target segment at 480×832 resolution, trained for 33,000 steps. §C
• The geometry-condition ablation compares full surface normals against no geometry, a textured render and a sparse turntable from 25 to 9 views. §4.3
• A preference study with 44 participants on 24 unseen videos reports OmniShowcase chosen most often on all four criteria. §B