-
Notifications
You must be signed in to change notification settings - Fork 217
Style Guidelines
The Stylist agent applies visual formatting rules based on academic conference conventions. This page documents what those guidelines contain and how they influence output.
The style guidelines are loaded from data/guidelines/ and fed to the Stylist agent as part of its prompt context. They are based on NeurIPS formatting conventions, which are widely used across ML venues.
The guidelines specify a constrained color palette suitable for academic figures:
- Muted, distinguishable colors that work in both color and grayscale print
- Consistent use of color to encode meaning (e.g., same color for all instances of a component type)
- Avoidance of overly saturated or neon colors
- White or light backgrounds for readability
- Left-to-right or top-to-bottom flow for sequential pipelines
- Consistent spacing between components
- Hierarchical grouping using bounding boxes or background shading
- Sufficient whitespace to avoid visual clutter
- Sans-serif fonts for labels (consistent with conference templates)
- Font sizes that remain legible when the figure is scaled to single-column width (~3.25 inches)
- Bold for component names, regular weight for descriptions
- Consistent capitalization
- Consistent arrow styles (solid for data flow, dashed for optional or conditional paths)
- Clear directionality
- Labeled arrows where the connection semantics aren't obvious from context
- Every element should serve a communicative purpose
- Favor clarity over visual complexity
- The diagram should be interpretable without reading the caption, and the caption should add context without being required for basic comprehension
The guidelines are text files, not code. You can modify them to suit different venues or preferences:
data/guidelines/
└── neurips_style.txt # Default style guidelines
To use different guidelines, either modify the file directly or create a new file and point to it in your config. The Stylist agent reads whatever text is at the configured guidelines path and uses it as instruction context.
The Stylist refines the description, not the image. It can specify "use a muted blue for encoder blocks" but whether the Visualizer's image generation model follows that instruction depends on the model's capabilities. Gemini generally follows color and layout instructions well, but fine-grained typography control is inconsistent.
The guidelines are also optimized for methodology/architecture diagrams. Statistical plots generated via paperbanana plot go through a different code-generation path where Matplotlib handles styling more directly.