Thursday, 29 October 2026 · 6 min read
OmniShowcase generates 3D-consistent videos from multi-view images using geometry conditioning
OmniShowcase creates videos of a subject from four reference images by reconstructing 3D geometry as surface-normal maps and conditioning a video diffusion model on these plus the references, achieving state-of-the-art identity and 3D consistency on new benchmarks.
Based on the paper OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images by Junyeong Ahn, Geonwoo Kim, Kinam Kim, Gihwi Kim and 3 others
Challenges in generating consistent subject videos from limited views
Generating videos of a specific subject from a few reference images is difficult because unseen surfaces must be inferred as the camera viewpoint changes, often causing inconsistent geometry and appearance in generated videos.
Early subject-to-video (S2V) methods use a single reference image, which limits information about the subject’s full 3D structure and appearance, leading to errors when the camera reveals new sides.
Recent methods extend to multiple reference views to provide more appearance information, but they do not explicitly model the 3D geometry connecting these views, and training data often lack guaranteed 3D consistency across viewpoints.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Two key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short