EN
A video method uses 3D shape to keep subjects consistent
OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images · arXiv:2610.35684
In plain words
A video made from a few pictures may show sides of an object that none of those pictures reveal. This paper introduces OmniShowcase, a method that adds a reconstructed 3D shape to the pictures and uses training videos made from shared 3D objects.
In two evaluations, its videos scored better than four alternatives on measures of subject identity and consistency of shape across views. The tests cover particular collections of subjects, so the result does not mean every generated video keeps its subject perfectly consistent.
The research
When a generated video turns a subject away from the camera, it may reveal surfaces the reference images never showed. Can a generator keep the subject’s appearance and shape coherent across those new views? This paper tests whether adding explicit 3D shape information can help answer that question.
OmniShowcase combines 3D-rendered training videos with a reconstructed shape condition to help generated subject videos stay consistent across views.
The authors compared OmniShowcase with four alternatives, giving each four reference images. In tests of subject identity and shape consistency across changing views, they measured similarity to the references and consistency with other views or ground-truth shape.
OmniShowcase led the reported identity and 3D consistency measures on both the 50-subject ShowcaseBench collection and the 35-subject NAVI collection. That supports stronger measured consistency in these evaluations, while leaving performance on other subjects and videos untested.
These comparisons cover two tested subject collections, and the methods differ in how they generate videos; 3DreamBooth also optimizes for each subject.
The static geometry condition does not handle subjects that change pose across the reference views, and the paper reports that generated backgrounds can look synthetic.
What the evidence shows
- The unseen sides problem
OmniShowcase is a subject-to-video framework that takes reference images and a text prompt to generate a video. Its central challenge is that a moving camera can reveal surfaces absent from the references, leaving the generator to infer how they look and fit the subject’s 3D shape.
More reference views provide more appearance information, but do not by themselves specify one shared shape.
- Tie views to one shape
The framework addresses that gap in two parts. Its training pipeline renders a subject along a camera path from one 3D asset, so the subject’s appearance and shape remain tied together across views.
At generation time, it reconstructs geometry from the four reference images and supplies a turntable of surface-normal maps alongside those images. The turntable describes shape without requiring frame-by-frame alignment to the requested camera path.
- What the evaluations show
The authors compared OmniShowcase with four other methods, giving each four reference images. On ShowcaseBench, which tests 50 subjects whose 3D assets were not used for training, and on NAVI, with 35 subjects, they measured identity and 3D consistency alongside prompt alignment and video quality.
OmniShowcase led the reported identity and 3D consistency measures on both collections. These are measured results for those collections, not a guarantee about every video.
- Checking the geometry condition
A separate control comparison helps assess the geometry condition’s role. The authors trained variants with no geometry, a colored shape render or a sparser turntable, using the same initialization, backbone and training data, then evaluated them at the same training step.
The full surface-normal turntable scored better on several identity and 3D measures, though individual alternatives led on some auxiliary measures. The comparison supports a contribution from the condition in this setup.
Deep
- Original paper authors
- Junyeong Ahn, Geonwoo Kim, Kinam Kim, Gihwi Kim, Hoonjin Jung
- Submitted to arXiv
- License
- CC BY 4.0