Media Describe

Media Describe is the center of gravity of this pack - the node everything else orbits. Feed it a path to an image or video and it sends the media to a vision model (Gemini by default, or your local LM Studio) with a strict system prompt, then returns a structured description broken into five named paragraphs: subject, clothing, movement, scene, and visual style. Those five paragraphs are the whole design philosophy: separate concerns so you can override one without regenerating the rest.

The required input is a single string, media_processed_path - "path to the processed media file from Media Selection node." That's your cue for how the pack intends this to be fed: MediaSelection picks the media, optionally trims and resizes it, and hands you a clean path. You can also just type a path to an image or video file directly.

The optional inputs are where the power lives. llm_studio_options switches the backend from Gemini to your local LM Studio (the pack's local-first play - same output, no cloud, no key). overrides accepts an object from Media Describe - Overrides, letting you hand-tune individual paragraphs before they're assembled. use_custom_prompts flips it into manual mode so you can send your own custom_system_prompt and custom_user_prompt instead of the auto-generated ones - the defaults baked in are already fine (a Wan 2.2-optimized five-field JSON instruction set), but when you want total control, this is the switch.

Seven outputs, and you'll use most of them. all_media_describe_data is the full structured payload - the thing you feed Control Panel Overview to audit. raw_llm_json is the unprocessed model output. positive_prompt_json is the five-field structure as JSON (feed this to Control Panel Prompt Breakdown). positive_prompt is that same content flattened into one text prompt - the actual string you'd feed a CLIP text encoder or T2I model. prompt_request is a log of what was sent. And height/width are INTs - the inferred output dimensions (the node defaults to 832×480 for landscape, 480×832 for portrait), so your downstream resize/empty-latent nodes can stay in sync automatically.

Mechanically it's a Gemini call with retry logic (three attempts, five-second backoff for the ever-popular 429 overload) behind a media cache, so re-running with the same media can skip the API call. The prompt_style knob - set in GeminiUtilOptions or the LLM Studio options - switches between descriptive Text2Image output and single-sentence ImageEdit instructions, so the same node feeds both generation and editing workflows.

Installing

Part of Swiss Army Knife:

• ComfyUI Manager: search "Swiss Army Knife", install, restart.
• Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sammykumar/ComfyUI-SwissArmyKnife.git
cd ComfyUI-SwissArmyKnife
pip install -e .

The workflow it's built for

Media Selection → Media Describe → (Media Describe - Overrides →) Control Panel breakdown → wire positive_prompt into your Wan 2.2 T2I/VACE pipeline, with height/width driving your latent sizes. That's the pack's canonical loop, and it's a genuinely nice one: describe media once, turn it into a structured prompt, iterate on individual paragraphs instead of the whole thing.

Gotchas

• Gemini needs a key in the pack's ComfyUI settings (or GEMINI_API_KEY env). LM Studio needs a running server. No backend, no output.
• media_processed_path is a string, not a tensor. If you want to describe something already in your graph as an IMAGE, route it through MediaSelection or save/point at a file first.
• The retry logic retries, but it doesn't fix a bad prompt. If the five-field JSON comes back malformed or empty, check the model choice and the media path before assuming a transient API hiccup.

For a pack called Swiss Army Knife, this is the main blade. Everything else - selection, overrides, control panels, caption combiners - exists to make this one node's output more useful.