Reproduce “Interactive Image Generation Using Scene Graphs” 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
Gaurav Mittal, Shubham Agrawal, Anuva Agarwal, Sushant Mehta, Tanya Marwah
Mittal et al. introduce an interactive image generation method that enables users to progressively build complex visual scenes by incrementally adding objects and relationships via scene graphs. Their approach successfully preserves previously generated content across steps and can be trained effectively on real-world datasets, even without intermediate ground-truth images.
Key insights
- Integrating previously generated image content directly into the generator's input helps maintain visual consistency across incremental steps.
- Applying perceptual loss between intermediate generated images and the final ground truth enables training on real-world datasets without requiring explicit intermediate supervision.
- Distributing complex scene generation across multiple steps can enhance overall image quality by reducing the generation load on the model at each stage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.