多任务画布数据集
收藏资源简介:
本研究构建的多任务画布数据集是一个大规模以人为中心的跨帧图像集合,包含来自100万个独特身份的600万张图像。该数据集通过跨帧采样策略实现灵活的组合构建,支持空间画布、姿态画布和边界框画布三种变体的生成。数据集主要应用于多模态控制的组合图像生成领域,旨在解决扩散模型在同时处理文本提示、主题参考、空间排列和姿态约束时的协同控制难题,为数字艺术和内容创作提供精准的可控生成能力。
The multi-task canvas dataset constructed in this study is a large-scale human-centric cross-frame image collection, containing 6 million images from 1 million unique identities. This dataset enables flexible composite construction via a cross-frame sampling strategy, and supports the generation of three variants: spatial canvas, pose canvas, and bounding box canvas. The dataset is primarily applied in the field of multimodal-controlled composite image generation, aiming to address the collaborative control challenges faced by diffusion models when simultaneously processing text prompts, subject references, spatial arrangements, and pose constraints, thereby providing precise controllable generation capabilities for digital art and content creation.




