Answers are grounded in this paper and cite their sources.
Recent advances in video generation and growing demand for immersive content highlight the need for VR180-specific training data. However, training VR180 video generation models requires large-scale video-text pairs that reflect stable binocular correspondence and hemispherical projection characteristics, which existing datasets do not sufficiently provide. We introduce WebVR180, a large-scale, high-resolution video clip-caption dataset for training VR180 video generation models. Sourced from VR180 videos collected from the web, the dataset is refined through a staged multi-metric curation pipeline. This pipeline rigorously evaluates stereoscopic structural validity across five key metrics—depth range, color mismatch, geometric alignment error, sharpness mismatch, and left–right view correspondence anomalies—while further assessing peripheral vignetting, brightness distribution, text overlays, and aesthetic quality. The curated clips are annotated with dense captions describing scene content and temporal changes, resulting in 396,597 VR180 clip-caption pairs at 2048×1024 resolution. Curation-stage ablations show that the model trained on the fully curated data achieves the lowest errors across all reported stereoscopic mismatch metrics. Furthermore, fine-tuning large-scale video generation models with the monocular ERP180 representations of WebVR180 and evaluating perspective-reprojected outputs demonstrate its applicability across different models and practical viewing perspectives.
No takes yet. Share an insight, caveat, or question.
Choi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: