UniData introduces a universal instruction generation pipeline that converts brief user requirements into multi‑round, multimodal instructions. First, the system expands the requirement into several diverse events; then, an any‑to‑any large model generates multimodal instructions based on these events; finally, it refines the data by correcting irrelevant or redundant inference flows using correlations across instruction rounds, thereby improving data quality.
To train this pipeline, the authors built UniDataset, comprising 20,000 entries spanning nine modalities, which markedly boosts multimodal generation performance. Experiments demonstrate that UniData achieves state‑of‑the‑art instruction quality and can further enhance the understanding and generation capabilities of other multimodal models.
Review