Image generation and vision plugins: 3 directions

Image plugins are not one category: generation turns prompts into images, vision understands images, and attachments handle upload and storage. Using the wrong one produces "it cannot see my image" or "it does not generate."

Three directions

Separate by direction before installing.

Generation and vision are opposite. Generation outputs images; vision reads them. A generation plugin will not describe your screenshot.

Attachments are infrastructure. They handle upload, paths, and saving, and may coexist with generation or vision.

Confirm the model. Some models support only text; vision requires a vision-capable model, or the image is ignored.

Image generation

Prompt, output, and saving are the three things to confirm.

1. Confirm the plugin, provider, and key.
2. Write a clear prompt: subject, style, size.
3. Set output size and format.
4. Confirm where output is saved.
5. Review and retry; fix prompts before changing models.

Keys and billing are usually separate. Many generation plugins call third-party services with their own keys and cost (source).

Generation takes time. Large images and queued services may take tens of seconds; do not assume failure immediately.

Vision and image understanding

Vision plugs images into the model's understanding.

1. Confirm the model supports vision.
2. Upload images with a clear question.
3. Describe the focus: regions, text, or layout.
4. Ask for structure when reading screenshots: lists or tables.
5. For sensitive images, check privacy terms first.

Vision quality depends on resolution and content, and a vision plugin cannot replace a generation plugin. Small text, blur, and complex layouts reduce accuracy; crop and enlarge before uploading. Vision understands, not creates.

Attachments and storage

Attachment handling decides "where are the files."

1. Confirm the upload entry and size limits.
2. Confirm save location: session directory or a custom path.
3. Clean old attachments periodically.
4. Back up important images.
5. Do not put sensitive images in shared directories.

Storage location is often overlooked. Inline images may only exist in the session; exporting requires a separate step.

Cleanup prevents bloat. Attachments accumulate; a monthly cleanup is simpler than disk cleanup later.

Keys and cost

Three cost sources.

1. Generation: per image, affected by size and model.
2. Vision: per call, affected by image count and resolution.
3. Storage: disk, usually local and free.

Check quotas before batch generation. Batch jobs are the easiest way to exhaust quota; estimate and run in stages.

Keys belong in configuration, not prompts. Do not paste keys into conversations.

Privacy boundaries

Images can contain more private information than text.

1. Screenshots: may include accounts, tokens, and internal data; crop before uploading.
2. Identity documents: do not upload unless the service is trusted and necessary.
3. Third-party services: where images are sent is decided by the plugin and provider.
4. Retention: check whether services retain images and for how long.
5. Sharing: exported images may contain private context.

Privacy depends on the destination, not the plugin. Read the provider's terms; unknown destinations should not receive sensitive images.

Choice and combination

Choose by task, combine only when needed.

Confirm model capabilities first. Vision and generation models are often different; one plugin cannot cover both.

Start with one, then add. Adding both at once makes cost and failure attribution harder.

Next steps

Direction, keys, storage. For plugin categories, read What DSH plugin categories exist. For workbench plugins, read Editor and workbench plugins. For voice plugins, read Voice plugin types.

Sources: dsh-image-gen plugin page, DeepSeek Harness docs, deepseek-harness repository