Continuous and Gradual Style Changes of Graphic Designs with Generative Model
summary
Proposes a pixel-based encoder–decoder that generates layout segmentation images with dual critic networks (KL encoder constraint and Wasserstein decoder constraint), a 128-d latent space and linear interpolation for style morphing between six dataset categories. Trains on ~3,904 annotated layouts, converts segmentations to final designs via contour detection and fill, and reports better FID than a Wasserstein autoencoder baseline. A 47-worker study finds 72% prefer its variation richness and 74% prefer its visual appeal; many interpolated layouts are novel.
AI-written summary · about
related talks
01h 17mDeep generative models for image...
13m[Interactive] Neural Style Transfer: The Art and Science of Generative AI
17mProcedural Modeling Using Autoencoder Networks
15mDrag Makeovers, via AI!
40mLearn to be a painter using Neural Style Painting
25mMixing art with Python- an introduction to Style Transfer
41mBuilding Generative Image & Video models at Scale
58mStyle Transfer AI - How Neural Networks Generate Art