OmniShowcase: generating subject videos that stay consistent across a full 360 degrees from multi-view images

Problem

In S2V (Subject-to-Video), which creates a video of a specific subject from a few reference images, the subject's shape and appearance tend to break down when the camera shows a side not present in the input. Recent methods that use multi-view references also do not explicitly provide the 3D structure linking the views. In addition, training data produced with camera-controlled video generators has no guarantee that the same 3D subject is preserved as the viewpoint changes.

Method

(1) Automatic data synthesis: 3D meshes from Objaverse and other sources are rendered in Blender to produce a target video and four reference views at 90° intervals. Subject pixels are left as is, and only the background is generated with video inpainting to composite a realistic scene. There are 12 camera motion types and 88 trajectory configurations. (2) Geometric condition: the shape is reconstructed with 3D Gaussians from the four references and rendered while rotating in yaw to produce a 25-frame surface-normal turntable. The same procedure is used in training and inference. (3) A pretrained I2V diffusion transformer (Wan) is tuned with LoRA. Reference, normal and target latents are concatenated along the time axis, providing the condition without aligning normal and target frames one-to-one.

Results

• The authors built Showcase-4K, a training dataset of 4,000 videos, and ShowcaseBench, an evaluation set of 50 subjects whose 3D assets do not overlap with the training data.
• The authors report the best performance on both subject identity (DINO and CLIP similarity) and 3D consistency (MEt3R, GT-CD) on ShowcaseBench and NAVI (35 subjects), compared with Phantom-MV, MAGREF-MV, MV-S2V and 3DreamBooth.
• The authors report that video quality and motion remained competitive.
• Unlike 3DreamBooth, which requires per-subject optimization, it applies to new subjects without additional optimization.

Limitations

The provided text is cut off before the experimental figures, so the quantitative comparison numbers could not be checked. The training data consists of synthetic rendered videos at the 4K scale, and the subjects are mainly static objects. The geometric condition depends on the quality of the reconstruction from the four references. Some evaluation metrics rely on external models such as GroundingDINO and SAM masks and π³ reconstruction.

Why it matters

It shows that conditioning on 3D geometry is effective for generation tasks that need an accurate subject shape, such as videos that rotate a product through 360 degrees. The pipeline for building view-consistent training data from 3D assets is also worth reusing.

Key terms

S2V(Subject-to-Video)

A task that generates a video featuring a specific subject from reference images while preserving that subject's appearance.

Surface-normal turntable

A sequence of images rendered by rotating a reconstructed 3D shape in place and recording the direction (normal) of each surface. It is fed to the model as a cue about 3D structure.

LoRA

A technique that freezes the existing weights and trains only small low-rank matrices, so large models can be fine-tuned at low cost.

In-context conditioning

An approach that concatenates conditioning information (reference images, normals) to the front of the input sequence instead of using separate channels, so the model can refer to it through attention.

Original

• arXiv abstract
• PDF
• HTML full text
• Code
• Project
• Hugging Face

Authors: Junyeong Ahn, Geonwoo Kim, Kinam Kim, Gihwi Kim, Hoonjin Jung, Hyojin Jang, Jaegul Choo · Posted 2026-10-26 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.