Which cheap vision model would you recommend for ingesting category diagrams and producing mermaid facsimiles?