Describer

If you've ever stared at a vision model's output and thought "that's a description, but not the useful kind" - this node exists for you. The Describer is a small but sharp idea: instead of hand-writing a "what's in this image" prompt every time, you pick a type of description and get back a matched pair of prompts tuned for that job. It's a prompt kit for vision models, nothing more, and nothing less.

It ships with three describers, each backed by a JSON file in the pack:

• Camera Angle - extracts shot type, camera height, tilt, horizontal position, depth of field, and lens characteristics, and explicitly ignores the scene's content. The system prompt is ruthless about scope: "only the photographic and cinematographic properties of the shot."
• Interior - describes a room: layout, architecture, furniture, lighting, palette, style. And it's told to ignore any people entirely.
• Outfit - describes clothing and accessories in detail, and is told to not describe the person wearing them.

That disciplined scoping is the whole value. A generic vision prompt mixes subject, setting, and gear into one mush; these prompts fence off exactly one axis so the model stays on task.

How it works

The node reads the describers/ JSON folder at load time and populates the describer dropdown from whatever files are there - so the pack can grow describers without a code change. Pick one, and you get two outputs: user_prompt (the question, e.g. "Describe the camera angle of this shot.") and system_prompt (the long, strict role definition). It also registers a /studiodeep/describers API route so the frontend can fetch or reload the list live.

Outputs: user_prompt and system_prompt (both STRING).

The natural wiring is into fal.ai Vision - user_prompt into its prompt input, system_prompt into its system_prompt input. The Camera Angle describer plus a vision call is a fast way to reverse-engineer how an existing reference was shot, which then feeds straight back into T2V Prompt Builder as locked angle/shot/movement values.

Installing

ComfyUI Manager → "StudioDeep", or:

cd ComfyUI/custom_nodes
git clone https://github.com/studiodeep-ai/comfyui-studiodeep StudioDeep
pip install -r StudioDeep/requirements.txt

Restart. No models, no keys, nothing - the Describer is fully local until you connect the vision model.

Troubleshooting

• Dropdown only shows (none) - the describers/ folder wasn't found or is empty, usually from installing the pack into a weird path. Confirm StudioDeep/describers/*.json exists next to nodes/.
• Adding your own describer - drop a JSON file with label, user_prompt, and system_prompt keys into describers/ and restart (or hit the reload route the frontend uses). That's a genuinely nice extensibility seam, and it's cheap to lean on.
• "It described the person anyway" - the system prompt is strong but models are models. If an Interior description wanders, bump up the model in fal.ai Vision (Sonnet-class handles strict instruction-following better than Flash).

It's a small node with one job, and it does that job well.