StereoWorld: Geometry-Aware Monocular-to-Stereo Video Generation

TL;DR

We propose StereoWorld, the first fully end-to-end diffusion framework that adapts a pretrained monocular video generative model into a stereo generator with high visual fidelity and geometric accuracy.
We build a large-scale, high-definition stereo video dataset aligned with human-IPD, featuring over 11M curated Blu-ray SBS video frames across diverse genres with comprehensive evaluation metrics.
Extensive experiments demonstrate that StereoWorld substantially outperforms prior works in visual quality, geometric consistency, and temporal stability, with clear advantages in both objective metrics and subjective perception.

Abstract

The growing adoption of XR devices has fueled strong demand for high-quality stereo video, yet its production remains costly and artifact-prone. To address this challenge, we present StereoWorld, an end-to-end framework that repurposes a pretrained video generator for high-fidelity monocular-to-stereo video generation. Our framework jointly conditions the model on the monocular video input while explicitly supervising the generation with a geometry-aware regularization to ensure 3D structural fidelity. A spatio-temporal tiling scheme is further integrated to enable efficient, high-resolution synthesis. To enable large-scale training and evaluation, we curate a high-definition stereo video dataset containing over 11M frames aligned to natural human interpupillary distance (IPD). Extensive experiments demonstrate that StereoWorld substantially outperforms prior methods, generating stereo videos with superior visual fidelity and geometric consistency.

Method

StereoWorld-11M Dataset

Dataset	Domain	IPD-aligned	Available	Frames
Spring	Optical Flow	✗	✓	5K
Sintel	Optical Flow	✗	✓	1K
VKITTI2	Driving	✗	✓	21K
PLT-D3	Driving	✗	✓	3K
IRS	Robotics	✗	✓	103K
TartanAir	Robotics	✗	✓	306K
3D Movies	Movies	✓	✗	75K
StereoWorld-11M	Movies	✓	✓	11M

We curated a new dataset tailored for stereo video generation with baseline (distance between two cameras) aligned to natural human perception. We collected and cleaned over a hundred high-definition Blu-ray side-by-side (SBS) stereo movies spanning animation, realism, war, sci-fi, historical, and drama, ensuring visual diversity and richness.

All videos are unified into SBS by stretching and horizontal cropping to obtain left–right views, each at 1080p, 16:9, 24 fps. To match the base model requirements (480p resolution, 81-frame inputs), we uniformly downscale to 480p.

To enhance motion diversity and temporal density, we uniformly sample 81 frames per clip at fixed intervals.

Framework

Before training, we use Video Depth Anything and Stereo Any Video to obtain the depth maps D_r and disparity maps Disp_gt, and the left-view videos are then concatenated with the right-view videos and corresponding depth maps along the frame dimension in the latent space as conditioning inputs. During training, a lightweight differentiable stereo projector estimates the disparity between the input left-view and the generated right-view, which is supervised by disparity maps Disp_gt via disparity loss to enforce accurate geometric correspondence. Additionally, the last few DiT blocks are duplicated to form dual branches, allowing the model to learn RGB and depth distributions separately to further supplement geometric information. During inference, only the shared and RGB DiT blocks are used, taking the monocular video as the sole input.

Practical and Scalable Optimization

Temporal Tiling Strategy

During training, the first few frames of noisy latents are replaced with ground-truth frames with a probability p. During inference, long videos are split into overlapping segments, with the last frames of the previous segment used to guide the next, ensuring temporal consistency.

Spatial Tiling Strategy

During inference, high-resolution videos are encoded into latents, which are split into overlapping tiles. Each tile is denoised independently, and then the tiles are stitched back to the original size with overlapping regions fused before decoding.

Comparison (Generated Right-View 480p)

Method	PSNR ↑	SSIM ↑	LPIPS ↓	IQ-Score ↑	TF-Score ↑	EPE ↓	D1-all ↓
GenStereo	19.4486	0.6803	0.3008	0.4047	0.9642	35.0022	0.8954
SVG	18.0256	0.5881	0.3467	0.4714	0.9706	33.2508	0.9630
StereoCrafter	23.0372	0.6561	0.1869	0.4370	0.9685	24.7784	0.5271
Ours	25.9794	0.7964	0.0952	0.5019	0.9704	17.4527	0.4213

GenStereo

SVG

StereoCrafter

Ours

GT

00001_segment_0137

00036_segment_0019

00093_segment_0393

00093_segment_0455

00095_segment_0156

00127_segment_0191

00134_segment_1314

More Examples (Side-by-Side Format)

Example

Citation

@misc{xing2025stereoworldgeometryawaremonoculartostereovideo, title={StereoWorld: Geometry-Aware Monocular-to-Stereo Video Generation}, author={Ke Xing and Longfei Li and Yuyang Yin and Hanwen Liang and Guixun Luo and Chen Fang and Jue Wang and Konstantinos N. Plataniotis and Xiaojie Jin and Yao Zhao and Yunchao Wei}, year={2025}, eprint={2512.09363}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2512.09363}, }