A full 360-degree orbit video of a product now takes just 4 reference images. On Oct 26, a team led by KAIST AI, together with Inha University and FLIPTION, posted OmniShowcase to arXiv (2610.35684); three days later it ranked #2 on Hugging Face Daily Papers for Oct 29 with 42 upvotes. The problem it targets is precise: in subject-to-video (S2V) generation, once the camera swings to surfaces the references never showed, the model starts inventing the subject.
One reference image cannot carry 360 degrees
The standard subject-to-video recipe conditions on a single reference image, forcing the video model to infer the subject's unseen appearance as the viewpoint changes. Recent methods extend this to multiple reference views, which adds information but never tells the model explicitly what 3D structure connects those observations. The paper points to a subtler flaw in training data itself: existing corpora do not guarantee that viewpoint changes preserve the same underlying 3D subject — the model is effectively studying a course with no answer key.
Two weapons: geometry-built data, normals as a turntable
The first weapon is a fully automated data synthesis pipeline. Starting from one curated 3D asset, Blender renders both the target orbit video and four reference views, and the target is composited into a prompt-driven scene. The key detail: only the background is inpainted, so the subject's far side always comes from real rendering and is never hallucinated. This yields Showcase-4K, a training corpus covering twelve camera motions across 88 trajectory configurations, about 4,000 videos, each with reconstruction-derived surface-normal conditioning. A held-out ShowcaseBench ships the source meshes and camera parameters behind every video — an evaluation set that carries its own answer key, which is rare in video generation.
The second weapon is the geometry conditioning: the four reference views are reconstructed in 3D and rendered as a turntable of surface-normal maps, then fed into the reference pathway alongside the references. The base model is a Wan2.1 image-to-video diffusion transformer with lightweight LoRA tuning; the normal turntable requires no frame-wise alignment with the generated video and no per-subject optimization. The turntable itself is a fixed yaw sweep, but the generated camera is not bound to it — the target follows whichever trajectory the prompt names, from 360-degree orbits and helix descents to trucks, arcs and dolly moves.
Numbers: geometric error halved, aesthetics not overclaimed
On its own ShowcaseBench, OmniShowcase scores DINO 0.799 and CLIP 0.925 for identity and GT-CD 0.0824 for geometry, leading all identity- and geometry-related metrics against four multi-view baselines: Phantom-MV, MAGREF-MV, MV-S2V and 3DreamBooth. Versus the strongest of them, geometric error drops about 52% on the held-out set and about 46% on the real-world NAVI objects — the "halved" phrasing comes from the authors' own paper-page note and checks out against the quantitative tables on the project page. Worth noting: its aesthetic score of 0.574 sits below MAGREF-MV's 0.614, so geometry is a hard win while visual appeal only ties. The ablation is equally clean: removing the normals, swapping in a textured render, or thinning the turntable degrades MEt3R from 0.190 to around 0.26, confirming the normal condition as the source of the gains.
Cold water: paper first, all four artifacts unreleased
The GitHub repository was created on Oct 29 with a single initial commit and 8 stars. On the README release checklist, all four items — Showcase-4K, ShowcaseBench, inference code, training code — remain unchecked, and the Hugging Face page shows zero models citing the paper. In other words, the paper and the demos come first; the full reproducible artifact set is still on the way. Teams waiting to train 3D consistency on the dataset can do little right now except star the repo.
For the industry, the direction matters more than any single benchmark: the next battleground in video generation may not be bigger models, but who can turn 3D geometry into training data and conditioning signals. The normal-turntable recipe is cheap — LoRA-level tuning — yet it directly kills the old "hallucinate the far side" failure. As product showcases and digital-asset workflows start demanding 360-degree consistency, whoever solves geometry first gets the ticket.
References: arXiv:2610.35684 (https://arxiv.org/abs/2610.35684); project page justin4ai.github.io/OmniShowcase; GitHub github.com/justin4ai/OmniShowcase