LM Studio plugin for image metadata extraction, visual analysis, and object detection powered by a Qwen3-VL model served through LM Studio's Vision API.
| Tool | What it does |
|---|---|
analyse_image | Extracts generation metadata from PNG files; optionally describes visual content |
detect_object | Object detection with labeled bounding boxes drawn on the result image |
annotate_image | Highlights detected objects or redraws stored detections with adjusted boxes |
The plugin uses the configured LM Studio-compatible API server for vision inference and does not start a local vision sidecar.
Standalone vs. draw-things-chat
This plugin works standalone — any LM Studio agent can call
analyse_image,annotate_image, anddetect_objectdirectly. To enable the agent to "see", evaluate, and autonomously correct generated images, combine it withdraw-things-chat. Vision Promotion — an exclusive feature ofdraw-things-chat— automatically routes generated images to the agent for review. Without it, generated images are displayed to the user but the agent itself cannot inspect them.
analyse_imageUse-case — generation metadata extraction (primary):
Reads prompt, model, seed, steps, and all other parameters embedded in PNG files produced by Draw Things and compatible tools. The agent cannot do this on its own — the metadata is binary-encoded in the file. No vision model is invoked for this; pass no prompt parameter.
Use-case — visual description (secondary, on demand):
Natural-language image description via Qwen3-VL. Only invoked when you pass a prompt parameter or configure a default Vision Prompt. Useful for captioning, evaluation, or quality checks.
Accepts attachments (a1), generated variants (v2), indexed images (i3), pictures (p1).
detect_objectannotate_imageThe plugin is platform-independent as long as it can reach an LM Studio-compatible server that exposes the required Vision API endpoints.
analyse_image metadata extraction works without any model. Visual descriptions, annotations, and object detection require a Qwen3-VL vision model available through the configured API server.
Start LM Studio Server and make sure the configured server can load the Qwen3-VL model key used by the plugin.
The plugin calls LM Studio's internal Vision API endpoints:
/api/v1/models/api/v1/models/load/api/v1/chatConfigure a Qwen3-VL model key, for example:
Use the LM Studio model key, not a filesystem path. The plugin asks LM Studio to load the model through /api/v1/models/load when needed.
- Models other than those specified above are not recommended — unless they're variants (finetunes, quantizations) of the listed standard models.
Download your preferred VLM (default: qwen/qwen3-vl-8b) to your server machine.
Open LM Studio Server and make sure it is reachable from the plugin machine. The default setting is:
https://lmstudio.ai/ceveyne/analyse-image
Activate it in LM Studio.
Open the plugin settings and set:
| Setting | Default | Recommended | Value |
|---|---|---|---|
| Vision API Base URL | http://127.0.0.1:1234/v1 | http://127.0.0.1:1234/v1 | LM Studio server URL. Existing /v1 or /api/v1 suffixes are normalized away before internal /api/v1 calls. |
| Vision API Key | (empty) | (API key if the server requires it) | Optional API key if the server requires it |
| Qwen3-VL Model | qwen/qwen3-vl-8b | qwen/qwen3-vl-8b | LM Studio model key, not a filesystem path |
The plugin loads/checks the configured Qwen3-VL model lazily on the first visual-analysis or detection request. PNG metadata extraction works immediately and does not invoke the Vision API.
detect_object bounding boxes from this plugin directly| Setting | Default | Description |
|---|---|---|
| Vision API Base URL | http://127.0.0.1:1234/v1 | LM Studio server URL. Existing /v1 or /api/v1 suffixes are normalized away; Vision tools call LM Studio internal /api/v1 endpoints. |
| Vision API Key | (empty) | Optional key for the API backend |
| Qwen3-VL Model | qwen/qwen3-vl-8b | LM Studio model key for Qwen3-VL, for example qwen/qwen3-vl-8b |
| Vision Prompt | (intentionally empty) | Default prompt for analyse_image visual descriptions. Leave empty to disable automatic visual description — metadata extraction is always active regardless. |
| Qwen3-VL: Object Detection Prompt | (built-in) | Detection instruction for annotate_image and detect_object. Leave empty to use the built-in hierarchical prompt. |
| Include Generation Metadata | on | Append Draw Things generation parameters (prompt, model, seed, ...) embedded in PNG files to each analyse_image result |
Target notation: a1 (attachment 1), v2 (variant 2), i3 (indexed image 3), p1 (picture 1).
See CHANGELOG.md for version history.
MIT
iN) with colored bounding boxes, plus a JSON summary with labels, coordinates, and crop percentages'all faces and hands', 'the dog', 'cars only'crop, inpaint, outpaint, and zoom-in from the process-image pluginqwen/qwen3-vl-8b
http://127.0.0.1:1234/v1
"Describe what's in attachment a1"
"Analyse image variants v2 and v3"
"Annotate the face and hands in a1"
"Detect all objects in image i1"
"What are the generation settings for picture p1?"
"Detect all faces and hands in a1"