ICLR2025
HQ-Edit: A High-Quality Dataset for Instruction-based Image Editing
Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Cihang Xie, Yuyin Zhou
摘要
a) Zoom in on the fox and add snowflakes falling around it. (b) Alter her hair color to black. (d) Replace the silver teapot with a ceramic blue and white patterned teapot with a similar wooden handle and a matching ceramic lid. (c) Replace the Amazon rainforest background with the underwater scenery of the Great Barrier Reef and adjust the parrot's position and wings to depict it flying. (e) Comparison between InstrictPix2Pix, HIVE, MagicBrush and HQ-Edit. Alignment Coherence Resolution HQ-Edit MagicBrush HIVE InstructPix2Pix Fig. 1: (a) -(d): example images and edit instructions from HQ-Edit. (e): we compare the dataset quality between our HQ-Edit and existing ones. Note that "Alignment" and "Coherence" are our newly developed metrics (introduced in Sec. 3.4) for measuring image/text qualities.