stdKonjac/DIM-Edit

Draw-In-Mind (DIM) 数据集旨在解决统一模型在文本到图像生成中的局限性，特别是在精确图像编辑方面。数据集由两部分组成：DIM-T2I 和 DIM-Edit。DIM-T2I 包含 1400 万个长上下文图像-文本对，以增强指令理解；DIM-Edit 包含 23.3 万个来自 GPT-4o 的思维链想象，作为图像编辑的明确设计蓝图。README 还描述了 DIM-4.6B-Edit 模型在各种图像编辑基准测试中的性能，并提供了如何使用数据集和模型的说明。The Draw-In-Mind (DIM) dataset is designed to address the limitations of unified models in text-to-image generation, particularly in precise image editing. The dataset is composed of two parts: DIM-T2I and DIM-Edit. DIM-T2I contains 14 million long-context image-text pairs to enhance instruction comprehension, while DIM-Edit comprises 233,000 chain-of-thought imaginations from GPT-4o that serve as explicit design blueprints for image editing. The README also describes the performance of the DIM-4.6B-Edit model on various image editing benchmarks and provides instructions on how to use the dataset and models.