DSH plugins for Image Generation

Image plugins add generation and editing to a session: text to image, image to image, asset variations for a UI, and shared history across devices. Most are thin adapters over an OpenAI-compatible endpoint, so the differences are in prompt handling, history, workspace writes and cost control — not in “the model can draw.”

Check where keys are stored, whether generated assets are saved into the workspace, and whether the plugin records the prompt alongside the file so an asset can be reproduced later. An image that cannot be regenerated from a stored prompt is a one-off, not a workflow.

How to evaluate

Four questions separate a useful plugin from a demo. Provider: which base URL and key does it call, and can you point it at your own endpoint? Persistence: are outputs written under an allowed workspace root with stable names? History: can you see prior generations and their prompts inside the harness, or only in a vendor dashboard? Cost controls: is there any cap, size limit, or model picker that keeps a long session from burning a key silently?

Design-oriented plugins go further — templates, editing surfaces, slide or layout export — and should be judged on whether those extras stay reviewable files rather than opaque project formats you cannot open without their UI.

What good looks like

A strong image plugin treats generation like any other tool: explicit inputs, files on disk you can open, and a history entry that names the prompt, model and seed when available. Multi-step editing keeps intermediate assets. Secrets live in config, not in the prompt. If the plugin also handles design systems or presentations, it should still write ordinary files into the workspace when you ask for an export.

Two polished listings cover this use case today. Depth of the tool surface matters more than the marketing phrase “AI image generation.”

Common failure modes

The usual failure is a key pasted into chat or into a settings card the model can read. The second is regenerating into an ephemeral path you cannot find tomorrow. The third is trusting a free public image backend for anything private: prompts and reference images leave your machine, and the plugin’s README is the only place that may say so.

Reproducibility is the real bar

An image plugin earns a permanent place in a workspace only when a colleague can rebuild the same asset from files on disk. That means prompt, model identifier, and seed (when the provider supports it) stored next to the output — not only in a vendor history page. Without that, you have a sketch tool, not a pipeline.

When a plugin also exports slides or design formats, open the export without the plugin. If you cannot, you have outsourced your assets to a black box.