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

Junyeong Ahn; Geonwoo Kim; Kinam Kim; Gihwi Kim; Hoonjin Jung; Hyojin Jang; Jaegul Choo

cs.CVOct 26, 2026

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

Authors: Junyeong Ahn, Geonwoo Kim, Kinam Kim, Gihwi Kim, Hoonjin Jung, Hyojin Jang, Jaegul Choo

Organizations: KAIST AI · Inha University · FLIPTION

Abstract

Generating videos of a specific subject from a few reference images remains challenging when the camera reveals surfaces that are not visible in the input. Early subject-to-video (S2V) methods condition on a single reference image and must therefore infer the subject's unseen appearance and geometry as the viewpoint changes. To address this limitation, recent methods extend S2V to multiple reference views, providing more information about the subject from different viewpoints. However, they do not explicitly specify the 3D structure that connects these observations. In addition, existing training data do not guarantee that viewpoint changes preserve the same underlying 3D subject. We present OmniShowcase, built on two components: (i) a fully automated data synthesis pipeline that produces training videos along diverse camera trajectories, with viewpoint variation driven by underlying subject geometry, even spanning full 360-degree orbits; and (ii) a geometry-conditioning method that provides the video model with a turntable of surface-normal maps reconstructed from the reference views without requiring temporal alignment with the generated frames. We construct Showcase-4K, a corpus of 3D-asset-based videos covering twelve camera motions across 88 trajectory configurations with reconstruction-derived surface-normal conditioning, and ShowcaseBench, a held-out evaluation set with the source meshes and camera parameters used to render each video. Experiments show state-of-the-art subject identity and 3D consistency while maintaining competitive video quality and motion. Project page: justin4ai.github.io/OmniShowcase

Figures & tables

Figure 1: OmniShowcase produces subject-driven videos that stay 3D-consistent across every revealed viewpoint.

Method	Data	References	General.	Geom.	Supervision
Phantom ( Liu et al., 2025 )	1M	single view	✓	$\times$	real video
MAGREF ( Deng et al., 2026 )	large	single view	✓	$\times$	real video
MV-S2V ( Song et al., 2026 )	14K	multi-view (4)	✓	$\times$	generated (87%) + real (13%)
3DreamBooth ( Ko et al., 2026 )	200K	multi-view (30)	$\times$	$\times$	static multi-view
OmniShowcase (ours)	4K	multi-view (4)	✓	✓	generated (3D asset render)

Table 1: Comparison of subject-to-video methods. Data : representative training-set order of magnitude. General. : generalization without optimization for inference. Geom.: explicit geometric conditioning. Supervision : source of the training signal.

Figure 2: Automated 3D-consistent data synthesis pipeline. From a curated 3D asset ( Deitke et al., 2023 ) , Blender renders the target orbit and four reference views. (a) The target is composited into a prompt-driven scene by video inpainting ( Jiang et al., 2025 ) ; (b) the four views are reconstructed in 3D ( Li et al., 2026 ) and rendered as the surface-normal turntable used as the condition. Every stage runs without human intervention.

Figure 3: Qualitative comparison of training data. Uni3C ( Cao et al., 2025 ) is an image-to-video model, so every surface the camera path uncovers is generated. Ours renders the subject from a 3D asset and generates only the background.

Figure 4: Overview of OmniShowcase. From $N{=}4$ reference views, we reconstruct 3D geometry and render a surface-normal turntable. The reference images and normal condition are injected into the reference pathway of a LoRA-tuned image-to-video diffusion transformer. The model uses appearance information from the reference views and 3D structure from the normal condition to generate the target video without per-subject optimization.

Figure 5: Qualitative comparison with baselines. 3DB denotes 3DreamBooth. Given a teapot and a jar of peach gummies, we compare OmniShowcase against four baselines. Red boxes mark where identity or 3D structure breaks down.

		Identity		3D consistency			Text alignment	Video quality
Set	Method	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$	ViCLIP $\uparrow$	Imaging $\uparrow$	Aesthetic $\uparrow$	Dynamic $\uparrow$
Held-out	Phantom-MV	0.621	0.871	0.444	0.2843	0.906	0.236	72.52	0.595	0.460
	MAGREF-MV	0.616	0.876	0.409	0.1789	1.177	0.242	73.88	0.614	0.760
	MV-S2V	0.688	0.896	0.351	0.2306	0.787	0.244	73.21	0.590	0.960
	3DreamBooth	0.707	0.907	0.325	0.1717	0.765	0.238	70.67	0.598	1.000
	OmniShowcase (ours)	0.799	0.925	0.190	0.0824	0.665	0.259	74.29	0.574	1.000

Table 2: Main quantitative comparison. Results on ShowcaseBench with

50

subjects and on NAVI ( Jampani et al., 2023 ) with

35

subjects. We evaluate identity preservation, 3D consistency, text alignment and video quality through nine metrics. Best results are in bold and second best are underlined .

Figure 6: Conditioning ablation. Given a biplane and a bulldozer, we compare the three conditioning variants against ours. Marked regions highlight where identity or 3D structure breaks down.

		Identity		3D consistency			Text	Video quality
Set	Geometry condition	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$	ViCLIP $\uparrow$	Imaging $\uparrow$	Aesthetic $\uparrow$
Held-out	Normals (ours)	0.799	0.925	0.190	0.0824	0.665	0.259	74.29	0.574
	None	0.752	0.903	0.265	0.2092	0.727	0.248	69.85	0.547
	Textured	0.755	0.910	0.270	0.1698	0.747	0.236	73.41	0.564
	Sparse	0.752	0.912	0.263	0.1777	0.698	0.250	71.84	0.552
NAVI	Normals (ours)	0.828	0.918	0.249	0.0975	1.074	0.260	74.58	0.551

Table 3: Geometry-condition ablation. None keeps only the reference views, Textured replaces the normals with a colored render of the same reconstruction, and Sparse thins the turntable from

25

9

views. Reference count is varied separately in Table . Best results are in bold and second best are underlined .

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Additional qualitative comparison. Four further subjects under diverse camera trajectories, with the same method ordering as Figure . Each block pairs the four reference views and the prompt, whose + [tag] suffix names the trajectory (Appendix ), with generated frames from every method.

Figure 8: Trajectory and scene control. One subject driven by several camera configurations, including arc, zoom-in and truck, with the background also varied by the prompt for each trajectory.

Figure 9: Trajectory control in a fixed scene. The same subject and background under different camera configurations, from full orbits to an arc and a zoom-out, where only the camera trajectory changes between rows. Inference only.

Figure 10: 3D reconstruction from generated video. The subject is segmented out of each method’s generated orbit and lifted to a coloured point cloud, the same reconstruction evaluated by GT-CD, shown here from four viewpoints so the geometric differences are visible directly. Baselines recover a thin shell of the surfaces their orbit actually sweeps, while ours closes the volume.

Figure 11: User study results. OmniShowcase is chosen most often on all four criteria, ahead of every baseline by a wide margin.

	ShowcaseBench						NAVI
#refs	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$	ViCLIP $\uparrow$	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$	ViCLIP $\uparrow$
$n=1$	0.702	0.902	0.301	0.1704	0.691	0.258	0.811	0.901	0.307	0.1875	1.091	0.245
$n=2$	0.768	0.918	0.241	0.1658	0.685	0.258	0.814	0.913	0.262	0.1915	1.075	0.249
$n=3$	0.783	0.921	0.206	0.1480	0.670	0.256	0.823	0.911	0.255	0.2050	1.074	0.249
$n=4$ (full)	0.799	0.925	0.190	0.0824	0.665	0.259	0.828	0.918	0.249	0.0975	1.074	0.260

Appendix

Table 4: Reference-count flexibility. Although OmniShowcase assumes four reference views, the model generalizes to fewer, while the full four-view setting performs best overall. The condition slots are filled by cycling the available views, and the surface-normal turntable is re-lifted from exactly those

n

views, so nothing about the missing views leaks in. Every row uses the same checkpoint and the same evaluation protocol as Table , and

n\!=\!4

repeats the OmniShowcase row of that table.

Source	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$
Uni3C ( Cao et al., 2025 )	0.7167	0.9191	0.3052	0.1409	0.6661
Ours (3D-asset render)	0.8030	0.9385	0.1871	0.1323	0.6318

Appendix

Table 5: Training clips scored as generations. Paired clips of the same subject under the same camera path and ground-truth geometry. Best results are in bold .

	Held-out					NAVI
Input	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$
Single view (original)	0.5921	0.8609	0.4736	0.1651	0.9254	0.6483	0.8677	0.5058	0.1791	1.4705
Four views (reported)	0.621	0.871	0.444	0.2843	0.906	0.690	0.882	0.462	0.2014	1.458

Appendix

Table 6: Phantom with one reference view and with four. We compare single-view input with the four-view setting used in Table . GT-CD uses ground-truth geometry. Best results per benchmark are in bold .

	Held-out					NAVI
Input	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$	DINO $\uparrow$	CLIP $\uparrow$	MEt3R $\downarrow$	GT-CD $\downarrow$	RotCov $\downarrow$
Single view (original)	0.5804	0.865	0.4538	0.1752	1.1974	0.657	0.8729	0.4529	0.1704	1.2712
Four views (reported)	0.616	0.876	0.409	0.1789	1.177	0.665	0.892	0.457	0.2145	1.284

Appendix

Table 7: MAGREF with one reference view and with four. We compare single-view input with the four-view setting used in Table . GT-CD uses ground-truth geometry. Best results per benchmark are in bold .

Figure 12: Camera trajectories. Sixteen representative trajectories out of the $88$ the vocabulary defines, one per tag family. Each panel draws the camera frusta sampled along the path, with the arrow giving the direction of travel and the grey dot marking the subject.

Tag	Sentence
Orbit	The camera circles [amount] around the object.
Helix-Up	The camera circles [amount] around the object while rising.
Helix-Down	The camera circles [amount] around the object while descending.
Helix-In	The camera circles [amount] around the object while moving closer.
Arc-Up	The camera circles partway around the object, rising from eye level to look down on it while moving closer, ending in front
Arc-Down	The camera starts high above the object and comes down to eye level while circling partway around and moving closer, ending in front

Appendix

Table 8: Motion vocabulary. The sentence each motion contributes to the prompt. For Orbit and Helix the amount is filled from Table . Every other motion is followed by a framing fragment from Table .

Suffix	Sentence fragment
90°	…circles a quarter of the way around the object
180°	…circles halfway around the object
270°	…circles three-quarters of the way around the object
360°	…circles all the way around the object

Appendix

Table 9: Rotation amount. Arc motions carry their rotation in the trajectory definition and take no suffix. Truck, Dolly and Zoom do not rotate.

Suffix	Sentence fragment
L $\rightarrow$ L	with the object held on the left of the frame throughout
L $\rightarrow$ C	with the object starting on the left of the frame and settling into the center
L $\rightarrow$ R	with the object starting on the left of the frame and drifting across to the right
C $\rightarrow$ L	with the object starting at the center of the frame and drifting across to the left
C $\rightarrow$ C	with the object held at the center of the frame throughout
C $\rightarrow$ R	with the object starting at the center of the frame and drifting across to the right

Appendix

Table 10: Framing transition. Where the object starts and ends inside the frame, with L, C and R for the left, center and right. The same fragment is appended regardless of the motion.

Figure 13: Failure cases. Left two. Prompt adherence and the commanded rotation both hold, but the surroundings render with flat shading and little material detail, so the clip reads as slightly synthetic. Right two. Both receive four reference views of a subject caught in four different poses. The upper one is prompted in the usual static style and settles into an averaged pose that matches none of the references, while the lower one is asked to follow the motion in step with the orbit and tangles in 3D at the transitions between poses.

Stage	Model	Input	Output
G1 asset gate	Gemini 2.5 Flash	4 rendered views, category label	6 binary criteria
G2 target gate	Gemini 2.5 Flash	4 reference views, 16 target frames	6 binary criteria
G3 geometry gate	Gemini 2.5 Flash	4 reference views, 16 RGB reconstruction frames	4 binary criteria
Category descriptions	Claude Opus 5	category label	appearance sentence, coarse type
Scene vocabulary	Claude Opus 5	—	scene bank (2,000 locations)
Captioning	Tarsier2-7B	video	condition and target captions

Appendix

Table 11: Where a language or vision-language model is used. Each gate accepts a sample only if all of its criteria score

1

. Gate calls use temperature

0

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 364 papers

3DreamBooth: High-Fidelity 3D Subject-Driven Video Generation Model
Mar 19, 2026Hyun-kyu Ko, Jihyeon Park, Younghyun Kim +23D-Aware Video Generation Controllable Video Generation
Yonsei University · Sungkyunkwan University
Tri-Prompting: Video Diffusion with Unified Control over Scene, Subject, and Motion
Mar 16, 2026Zhenghong Zhou, Xiaohang Zhan, Zhiqin Chen +83D-Aware Video Generation Video Generation
Adobe Research · University of Rochester
Track2View: 4D-Consistent Camera-Controlled Video Generation via Paired 3D Point Tracks
Jun 14, 2026Feng Qiao, Zhaochong An, Zhexiao Xiong +2Multi-View Consistency Camera-Controlled Video Generation
Washington University in St. Louis · University of Copenhagen

Mar 19, 2026J/K move · Enter open · S save

3DreamBooth: High-Fidelity 3D Subject-Driven Video Generation Model

Hyun-kyu Ko, Jihyeon Park, Younghyun Kim, Dongheok Park+1

Yonsei University · Sungkyunkwan University

3D-Aware Video Generation Controllable Video Generation Video Diffusion Models Multi-View Video Generation Temporal Consistency in Video Generation

Open paper AskPDF arXiv

Creating dynamic, view-consistent videos of customized subjects is highly sought after for a wide range of emerging applications, including immersive VR/AR, virtual production, and next-generation e-commerce. However, despite rapid progress in subject-driven video generation, existing methods predominantly treat subjects as 2D entities, focusing on transferring identity through single-view visual features or textual prompts. Because real-world subjects are inherently 3D, applying these 2D-centric approaches to 3D object customization reveals a fundamental limitation: they lack the comprehensive spatial priors necessary to reconstruct the 3D geometry. Consequently, when synthesizing novel views, they must rely on generating plausible but arbitrary details for unseen regions, rather than preserving the true 3D identity. Achieving genuine 3D-aware customization remains challenging due to the scarcity of multi-view video datasets. While one might attempt to fine-tune models on limited video sequences, this often leads to temporal overfitting. To resolve these issues, we introduce a novel framework for 3D-aware video customization, comprising 3DreamBooth and 3Dapter. 3DreamBooth decouples spatial geometry from temporal motion through a 1-frame optimization paradigm. By restricting updates to spatial representations, it effectively bakes a robust 3D prior into the model without the need for exhaustive video-based training. To enhance fine-grained textures and accelerate convergence, we incorporate 3Dapter, a visual conditioning module. Following single-view pre-training, 3Dapter undergoes multi-view joint optimization with the main generation branch via an asymmetrical conditioning strategy. This design allows the module to act as a dynamic selective router, querying view-specific geometric hints from a minimal reference set. Project page: https://ko-lani.github.io/3DreamBooth/

Abstract

Video Generation

Explore similar work

3DreamBooth: High-Fidelity 3D Subject-Driven Video Generation Model

Tri-Prompting: Video Diffusion with Unified Control over Scene, Subject, and Motion

Track2View: 4D-Consistent Camera-Controlled Video Generation via Paired 3D Point Tracks