People test Nano Banana with PDF paper to whiteboard. I did the exact opposite (quickchat.ai)

Instead of the usual demo—turning long PDFs into slick whiteboard diagrams—the author fed a photo of an office whiteboard to Nano Banana Pro (DeepMind’s Gemini-image) and asked it to generate the underlying PDF research paper. The model produced full LaTeX source and a coherent, high-level paper titled “A Unified Framework for Semantic Retrieval and Multi-Stage Ranking using Knowledge-Enhanced Context Chunks (ScCh).” The generated paper defines ScCh as a tuple (S, C, {A_i}) representing score vectors, dense context embeddings and knowledge-graph attributes, and outlines a Mapper → Spreader → Rescorer → Reducer/Aggregator pipeline that unifies dense retrieval with a “Large Hk” knowledge hypergraph, expands context via spreading, rescoring with a learnable model, and final aggregation. Technically, the demo shows vision-language models can infer structured, plausible technical artifacts (including LaTeX) from fuzzy visual cues: the model tried to incorporate visible equations and abbreviations into a full methodology using dense embeddings, knowledge-graph augmentation, normalization, and aggregation functions (max/average/learned sum). This is valuable for rapid ideation and sketch-to-paper prototyping, but highlights major hallucination and fidelity risks—outputs can be syntactically sound yet shallow or invented. The experiment is easy to reproduce via Gemini’s image interface and a prompt asking for LaTeX/PDF, underscoring both the creative potential and caution needed when using generative V&L models in research workflows.