Reproduce “StyLandGAN: A StyleGAN based Landscape Image Synthesis using Depth-map” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gunhee Lee, Jonghwa Yim, Chanran Kim, Minjae Kim
StyLandGAN, developed by Vision AI Lab at NCSOFT, is a StyleGAN-based framework that synthesizes landscape images using both depth maps and semantic segmentation maps as conditional inputs. This approach yields generated images with higher quality, demonstrated by a FID of 33.60, and improved depth accuracy, with an RMSE of 8.80, compared to previous methods, by leveraging depth's ability to convey structural information and providing a two-phase inference pipeline for user control.
Key insights
- Depth maps possess higher expressive power than semantic segmentation or edge maps, capable of encoding both linear (ridge) and planar (scale, texture) information vital for realistic and controllable landscape generation.
- A two-phase inference pipeline, separating 'structure selection' (depth map generation/editing) from 'style selection' (final image synthesis), offers a more intuitive and robust workflow for users to create diverse and structurally precise landscapes.
- Explicitly guiding image synthesis with depth information, even pseudo depth derived from monocular estimation, significantly enhances control over the generated image's 3D structure and improves overall perceptual quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.