Reproduce “FigGen: Text to Scientific Figure Generation” 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
Juan A Rodriguez, David Vazquez, Issam Laradji, Marco Pedersoli, Pau Rodriguez
FigGen, a latent diffusion model developed by researchers at ServiceNow Research and collaborators, introduces the problem of text-to-scientific figure generation. The model demonstrates the ability to capture high-level concepts from text, achieving an FID of 281.25 with its largest configuration (128-layer BERT encoder, CFG 5.0) when generating scientific figures.
Key insights
- Utilizing a custom-trained text encoder specifically for technical language significantly improves performance compared to general-purpose encoders for scientific figure generation.
- Classifier-Free Guidance effectively enhances the alignment between generated scientific figures and their textual prompts, with an optimal guidance scale identified (CFG 5.0).
- Latent diffusion models can learn and translate high-level semantic relationships from text into discrete graphical elements, demonstrating their applicability beyond continuous natural image generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.