moondream2 Caption

moondream2 Caption looks at an image and writes a sentence about it, entirely on your machine. No API, no key, no cloud round-trip - despite the product-sounding name, this is a tiny vision-language model living inside your ComfyUI install.

Reach for it when you want to caption a stack of images without a monster GPU. In the local-captioner pecking order, moondream2 sits a rung below JoyCaption and Florence-2 on quality, and comfortably above BLIP - which, as anyone who's trained a LoRA recently will tell you, has been "pure trash" for years. What moondream2 gives back for that quality gap is size: a couple of billion parameters, small enough to run on a laptop CPU if you don't mind waiting. For sorting a folder, finding a few good frames, or seeding a prompt, it's exactly the right tool. For building a serious training dataset, caption by hand or reach for JoyCaption instead.

How it works

The pack splits a VLM into two halves. You first run the image through moondream2 Encode, which runs the vision encoder and hands back a pre-encoded md2_image. This node takes that encoded result and runs the language half: it nudges the small LLM with a "describe this image" style prompt and streams out text. That two-stage design is why you can run the pricey vision step once and then caption or query the same image as many times as you like - worth remembering if you're asking more than one thing per render.

Inputs and output

The knobs you'll actually touch:

• model - the moondream2 you loaded with the DownLoad node.
• md2_image - the output of moondream2 Encode. There's no way around encoding first; this node won't accept a raw IMAGE.
• length - short gives you a one-liner, long a fuller description.
• max_tokens - the ceiling on generated text, 512 by default. Generous; you'll rarely touch it.

The caption output is a plain STRING. That's what makes the node flexible: wire it into a CLIPTextEncode text input to feed an image its own caption as the prompt for an img2img pass, or drop it into a Save Text node to build a .txt per image for dataset prep - the classic "load a folder, caption each image, save" loop.

Install

ComfyUI Manager → Custom Nodes Manager → search ComfyUI-moondream2, or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/zhilemann/ComfyUI-moondream2

Restart ComfyUI. The pack needs a reasonably recent transformers (more below) and pulls the model weights from HuggingFace on first use.

Common issues

The first-run trap: the DownLoad node is fetching a couple of gigabytes from HuggingFace the first time you run the graph. It looks hung. It isn't - let it finish once and it's cached forever after.

If you hit a load error that mentions transformers, that's the classic moondream failure mode: the model needs a newer transformers than ComfyUI's bundled environment ships. pip install -U transformers and restart usually clears it.

And manage expectations. At 2B this model writes fluent, mostly-accurate captions, but small models hallucinate small details, and the 0.5b variant is noticeably dumber. Great for batching and prompt reuse; verify anything you're about to train on.