Training vision-language models (VLMs) typically requires large-scale, high-quality image-text pairs, but collecting or synthesizing such data is costly. In contrast, text data is abundant and inexpensive, prompting the question: can high-quality multimodal training data be synthesized purely from text? To tackle this, we propose a cross-integrated three-stage multimodal data synthesis framework, which generates two datasets: Unicorn-1.2M and Unicorn-471K-Instruction. In Stage 1: Diverse Caption Data Synthesis, we construct 1.2M semantically diverse high-quality captions by expanding sparse caption seeds using large language models (LLMs). In Stage 2: Instruction-Tuning Data Generation, we further process 471K captions into multi-turn instruction-tuning tasks to support complex reasoning. Finally, in Stage 3: Modality Representation Transfer, these textual captions representations are transformed into visual representations, resulting in diverse synthetic image representations. This three-stage process enables us to construct Unicorn-1.2M for pretraining and Unicorn-471K-Instruction for instruction-tuning, without relying on real images. By eliminating the dependency on real images while maintaining data quality and diversity, our framework offers a cost-effective and scalable solution for VLMs training.

翻译：训练视觉语言模型通常需要大规模、高质量的图文对数据，但收集或合成此类数据成本高昂。相比之下，文本数据丰富且廉价，这引发了一个问题：能否仅从文本合成高质量的多模态训练数据？为解决此问题，我们提出一个交叉集成的三阶段多模态数据合成框架，生成了两个数据集：Unicorn-1.2M 和 Unicorn-471K-Instruction。在第一阶段：多样化字幕数据合成中，我们利用大语言模型扩展稀疏字幕种子，构建了120万条语义多样且高质量的字幕。在第二阶段：指令微调数据生成中，我们进一步将47.1万条字幕处理为多轮指令微调任务，以支持复杂推理。最后在第三阶段：模态表征迁移中，这些文本字幕表征被转换为视觉表征，从而生成多样化的合成图像表征。通过这一三阶段流程，我们无需依赖真实图像即可构建用于预训练的Unicorn-1.2M和用于指令微调的Unicorn-471K-Instruction。该框架通过消除对真实图像的依赖，同时保持数据质量和多样性，为视觉语言模型训练提供了一种经济高效且可扩展的解决方案。