What AI diagramming still gets wrong

First drafts became a feature

Napkin turns imported documents or selected prose into editable article visuals. Paid plans add SVG export, brand styles, and custom fonts. Its custom generation controls include a setting to stay closer to the supplied text.

Eraser returns diagram-as-code that can be edited manually or by later prompts. Mermaid AI adds conversational generation and visual editing to an established portable text format. Miro, Lucid, FigJam, Whimsical, and Excalidraw cover similar first-draft territory inside broader canvases. Piktochart extends the pattern to editable infographics generated from prompts, pasted text, reports, notes, or uploaded documents.

Transmute should not compete by promising another instant flowchart. Its narrower job is to keep a small explanatory visual source-controlled, inspectable, and ready to publish in both themes.

The artifact is the product

A flattened image hides its claims. A checked source exposes the labels, relationships, geometry, icons, and theme inputs that produced it. Review can ask whether an arrow was supported, why peer boxes differ in size, or which command produced the current export.

That continuity matters more as agents become the interface. A useful agent operation creates or patches a stable object. It should not regenerate the whole composition because one label changed. Eraser, Miro, Mermaid, Figma, QuiverAI, and tldraw now expose agent or MCP paths. tldraw's agent starter kit combines structured shape state with rendered visual context and validated actions.

Transmute keeps the headless renderer independent of an editor. A `.diagram.json` file produces light and dark SVG and PNG assets plus always-emitted editable `.tldr` interchange. Importing that file into tldraw Offline is an optional editing handoff; the editor is not required in CI or on a rendering server.

Images are becoming structured input

Two vector jobs are often described with the same word. Native vector generation reconstructs or invents semantic shapes. Conventional vectorization approximates supplied pixels with paths. The output may look similar, but the fidelity, licensing, and privacy contracts differ.

Adobe now makes the split visible inside one editor. Illustrator Image Trace converts raster pixels into editable vector graphics, while Concept to Vector uses prompts to produce a refined or stylized variation that may deliberately change the input.

Recraft prices native editable SVG generation separately from conventional vectorization. QuiverAI offers text-to-SVG and image-to-SVG behind an API and a hosted MCP server. Vectorizer.AI shows the mature developer surface for deterministic tracing: CLI, API, SDKs, previews, and several vector outputs.

Kittl and Canva integrate raster tracing with browser-based editing. Their advantage is continuity into a larger design suite; Transmute's test is whether local limits, measurable fidelity, and source-control provenance matter more for an agent workflow.

Transmute includes the local macOS and Linux form of that second job. The `vectorize` command uses a pinned VTracer engine, compares bounded trace profiles, rejects unsafe SVG, renders the candidate back to pixels, measures error, and emits a provenance receipt.

transmute image vectorize input.png --output input.svg --json

The command does not claim to recover lost design intent. It produces measured paths from caller-owned pixels. Remote semantic SVG providers can fit behind a later adapter without becoming a mandatory dependency.

Pixel similarity is not structural quality

Current vector research makes the distinction explicit. StarVector reports that pixel metrics alone can reward structurally poor SVG. VFIG adds structural-integrity evaluation for complex figures. Render-in-the-Loop uses intermediate renders and a render-and-verify pass to remove degenerate or redundant primitives.

Transmute needs both kinds of evidence. Raster-to-SVG conversion measures premultiplied color and alpha error, paths, bytes, and provenance. Diagram checks should also measure labels, nodes, relationships, crossings, connector runway, peer geometry, and light/dark contrast. A screenshot match cannot tell whether the diagram invented a hierarchy.

A small local compiler is the opening

Most direct products are proprietary cloud canvases with seats or credits. Open local diagram tools exist, but the reviewed set did not combine a restrained article grammar, prompt-boundary checks, representative icons, paired themes, Git source, optional editor handoff, and measured vector assets.

Transmute therefore optimizes for three to seven concepts, short labels, one primary reading direction, and no invented group, sequence, or emphasis. It is for the explanatory visual beside the work, not the infinite board around it.

The next tests

1. Semantic layout. The first stack source now owns ordinary horizontal or vertical sequences. Recreate the publication corpus before adding flow, grid, or a layout dependency.
2. Patchable agent tools. The first local server now exposes a fixed check, render, local vectorize, and hosted generate registry for a trusted local workspace. Compare it with the skill-only workflow before adding create, patch, or Mermaid interop.
3. Hosted demand. The authenticated free preview fixes model choice, retries, output type, and daily limits. Measure repeated use, provider cost, and abuse before enforcing the planned paid entitlement.
4. Grounded high-level diagrams. Map every node and edge to source evidence so “do not enrich the prompt” can be reviewed rather than trusted as an instruction.
5. Reusable vector assets. Admit normalized local SVG with required alt text, explicit recoloring, path budgets, and deterministic tldraw handoff.

Each proposal has a cheap failure test: existing diagram recreations, adversarial image fixtures, narrow-patch agent sessions, and round trips through supported formats. A general whiteboard, proprietary vector model, or production web editor comes later only if those tests reveal the need.

Method

Market sources were reviewed July 25, 2026; repository facts were refreshed July 26. The snapshot uses first-party product pages, documentation, pricing, public repositories, and primary papers. Vendor claims describe available products, not independent output-quality tests. Product proposals are hypotheses, not customer evidence.

The longer design argument lives in A visual grammar for AI-generated diagrams. The folded Graphics predecessor remains auditable on GitHub.