Suizi Huang
Yue Lin
Yinlin Ye
IEEE Transactions on Visualization and Computer Graphics (Proc. IEEE VIS), 2023
https://github.com/SerendipitysX/ChartSpark
https://www.youtube.com/watch?v=pBHWU3zjW6c
🟨In current work on pictorial visualization generation, the usual approach is to follow a retrieve and adapt pipeline. However, it is often difficult to find an image that perfectly aligns with the data trend.
🟨 Though current text-to-image models show promise for lay users to create a vivid picture, their training in natural images leads to suboptimal performance in generating visualizations.
🟨 We distilled design knowledge from 869 real-world examples, achieving two modes(Background + Foreground) through iterative fine-tuning with a cross-attention mechanism.
2023-12-18 21-15-57 (online-video-cutter.com).mp4
🟨 Users can switch the mask to personalize the object they want to generate.
🟨 A hybrid mode with conditional and unconditional, foreground and background can jointly contribute to a harmonious performance.
