Help us build a better reference dataset: submit papers with strong methodology diagrams

[dippatel1994]

Help us build a better reference dataset: submit papers with strong methodology diagrams

PaperBanana's output quality is directly tied to its reference examples. The current set works, but more diverse, high-quality samples across different diagram styles will meaningfully improve generation.

What we need

If you've read or written a paper with a clear, well-designed methodology diagram, submit it here. We're looking for papers where:

• The methodology diagram clearly illustrates system architecture, pipeline flow, or framework structure
• The diagram uses a landscape layout (roughly 1.5 to 2.5 width-to-height ratio)
• The methodology section in the paper text is self-contained enough to describe the approach without needing the full paper
• The diagram isn't just a results plot, ablation table, or data sample visualization

How to contribute

Option 1 (low effort): Drop a comment with the arXiv link and the figure number. We'll handle extraction and curation.

Option 2 (more involved): Submit a PR adding the parsed tuple (methodology text, diagram image, caption) to the reference set. See /data/README.md for the expected format.

What categories we're short on

The paper defines four diagram categories. We're currently weakest on:

• Science & Applications — domain-specific architectures outside core ML
• Vision & Perception — detection/segmentation pipelines, multimodal architectures

But any well-designed diagram from any category helps.

#Guidelines

• Only publicly available papers (arXiv, conference proceedings)
• One paper per comment for easier tracking
• If it's your own paper, even better — you can confirm we're extracting the right section

The generation pipeline is working. The problem is reference quality and diversity. This is the highest-leverage contribution anyone can make to the project right now.