cohort-rlwm/ByteMorph

ByteMorph-Bench数据集专注于基于指令的非刚性运动图像编辑。数据集分为五种编辑能力类型：相机缩放、相机运动、物体运动、人体运动和交互。数据集包含源图像和目标图像、标题以及编辑指令。主要用途是用于文本到图像和基于指令的图像编辑研究，目标用户是计算机视觉和AIGC领域的研究人员和爱好者。数据集结构包括编辑类型、图像ID、源图像、目标图像、编辑提示、编辑指令重写、源图像标题和目标图像标题等特征。数据集来源于Seaweed生成的原始视频，并采样成帧，由VLM进一步标注。

The ByteMorph-Bench dataset focuses on instruction-based image editing for non-rigid motions. The dataset is categorized into five types of editing capabilities: Camera Zoom, Camera Motion, Object Motion, Human Motion, and Interaction. It includes source and target images, captions, and editing instructions. The primary intended use is for research in text-to-image and instruction-based image editing, targeting researchers and hobbyists in computer vision and AIGC. The dataset structure features include edit_type, image_id, src_img, tgt_img, edit_prompt, edit_prompt_rewrite_instruction, src_img_caption, and tgt_img_caption. The dataset is derived from original videos generated by Seaweed and sampled into frames, with further captioning by VLM.