stdKonjac/DIM-T2I

DIM数据集包含两个互补部分：DIM-Edit和DIM-T2I。DIM-Edit是一个包含233,000个链式思维想象的数据集，由GPT-4o生成，用于提供明确的图像编辑设计蓝图。每个样本包括唯一标识符、源图像路径、目标图像路径和描述如何将源图像转换为目标图像的链式思维风格提示。该数据集旨在帮助统一多模态模型在图像编辑任务中重新平衡理解模块（设计师）和生成模块（画家）的职责。DIM-T2I是数百万个长上下文图像-文本对，用于增强模型的指令理解能力。数据集以JSONL格式提供，并遵循CC-BY-NC 4.0许可证。The DIM dataset consists of two complementary components: DIM-Edit and DIM-T2I. DIM-Edit is a dataset containing 233,000 chain-of-thought imagination samples generated by GPT-4o, which provides explicit image editing design blueprints. Each sample includes a unique identifier, source image path, target image path, and chain-of-thought style prompts that describe how to transform the source image into the target image. This dataset aims to help unify multimodal models by rebalancing the responsibilities of their understanding module (designer) and generation module (painter) in image editing tasks. DIM-T2I comprises millions of long-context image-text pairs designed to enhance models' instruction understanding capabilities. The dataset is provided in JSONL format and is released under the CC-BY-NC 4.0 license.