HP-Image-40K
收藏资源简介:
HP-Image-40K是由字节跳动等机构构建的大规模人-产品图像数据集,包含4万余条高质量样本,旨在解决广告和电商领域高保真图像生成的训练数据匮乏问题。该数据集通过预训练文本-图像模型合成初始样本,并经过自动化过滤流程(包括语义对齐、边缘分割、CLIP相似度筛选及文本一致性校验)确保数据多样性和细节真实性。其核心应用为支持基于参考图像的修复框架HiFi-Inpaint,通过高频特征增强和像素级监督,实现产品纹理、品牌标识等细粒度元素的高精度保留。
HP-Image-40K is a large-scale human-product image dataset constructed by ByteDance and other institutions, which contains over 40,000 high-quality samples. It aims to address the shortage of training data for high-fidelity image generation in the advertising and e-commerce domains. This dataset synthesizes initial samples using pre-trained text-image models, and then adopts an automated filtering pipeline including semantic alignment, edge segmentation, CLIP similarity screening and text consistency verification to ensure data diversity and authenticity of details. Its core application is to support the reference-image-based inpainting framework HiFi-Inpaint, which achieves high-precision retention of fine-grained elements such as product textures and brand logos through high-frequency feature enhancement and pixel-level supervision.
- 1HiFi-Inpaint: Towards High-Fidelity Reference-Based Inpainting for Generating Detail-Preserving Human-Product Images中国科学院大学; 香港中文大学; 字节跳动; 浙江大学; 德克萨斯大学奥斯汀分校 · 2026年



