Superside Florence-2 Detailed Caption

Sometimes you just want the caption and you don't want to think about which model, which task token, or which prompt format to type. This node is the fixed-purpose version of that: connect an image, run, and get a detailed natural-language caption back as a STRING. It runs Microsoft's Florence-2 Large on fal's hosted endpoint, so there's no prompt to write and no weights to download - the model picks the detailed-caption task itself.

That's the whole surface area, and it's refreshingly small: inputs are image and api_key, nothing else. Output is a single STRING, the caption. For captioning a batch of product shots into a text file, or seeding a prompt from a reference image before an img2img pass, that's genuinely all you need.

The background on Florence-2 is worth knowing because it's one of the community's favorite vision models for exactly this job. It's a tiny unified VLM (0.77B in the Large flavor, an order of magnitude smaller than the LLaVA-class captioners), MIT-licensed, and it does more than caption - one model handles captioning, detection, grounding, OCR and segmentation depending on the task token you feed it. It's also famously not prudish, which is why it became a training-caption staple. The KB's verdict on the model's limits carries over here unchanged: captions are detailed and faithful enough to describe but not to reproduce, and multi-subject attribution is its weak spot - two people in a frame and it will mix up who's wearing what. For a quick describe-the-image job that's fine; for a LoRA training set you care about, audit the output.

The one honest objection to this specific node: Florence-2 is an open-weight model that runs locally for free, and the KB documents it arriving in ComfyUI through Kijai's local node back in 2024. So why would you pay fal per call for it? The reasons are the usual API-node ones - no local install, no VRAM, it just works on a machine where you haven't set up the local path. If you're already running this pack and you have a key, the convenience is real. If you're captioning hundreds of images, the local route wins on cost alone, and it's one install away. This node is the "I just need one caption and I'm already here" option.

It's a fal call, so the standard pack rules apply: paste your key into the api_key widget (blank falls back to FAL_KEY, both blank errors immediately), calls are metered credits, and your image briefly leaves the machine. There's no sync_mode here and no queued-execution path - captions are fast, so it uses the plain synchronous call.

Install - ComfyUI Manager (search "comfyui-superside-nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes
pip install -r requirements.txt

Restart ComfyUI, find it under Superside. No model downloads - requirements.txt is fal-client, pillow, numpy, torch, requests. Community coverage of this pack is basically nil, so treat the README as the manual; it's current.