Fooocus Describe

Somewhere between "inpaint this" and "generate that" there's a much simpler workflow: look at an image and copy its vibe. That's this node. Fooocus Describe is the pack's image-captioner - feed it a picture, get a ready-to-use prompt string back, paste it into the Loader, regenerate something in the same spirit. It's the "describe image" feature from Fooocus, shipped as a single node.

What it does

It runs a vision-captioning model over the input image and writes out a natural-language description: subject, setting, lighting, style cues, the works. The output is designed to be prompt-shaped, not just descriptive - the whole point is that you can feed it straight back into the pipeline. Two flavours, chosen by image_type:

• Photo - default. Captions written like a photography prompt, tuned for realistic SDXL outputs.
• Anime - captioning aimed at anime/illustration style, producing the kind of description that works well with anime-style checkpoints.

It's an image-to-text node, so it has exactly one input that matters and one output:

• image - the reference image. Required.
• Output: a single STRING caption.

There are no weights or toggles. Either it's the kind of image you want a photo prompt for, or it's anime. That's the whole node.

Installing

Standard pack install - ComfyUI Manager, search "Fooocus Nodes", or clone the repo into custom_nodes and restart. Worth knowing: this is one of the pack's nodes with a hidden model download, a captioning model that pulls on first use. Give it a minute on the first run; it's cached after.

Where it fits

The cleanest use is a two-node loop: Describe on your reference image → feed the STRING into the Loader's positive field (or the Fooocus positive node) → generate variations that inherit the description. It pairs well with Fooocus Expansion if you want the caption embellished further before sampling, and with the Fooocus ImagePrompt if you want both the description and the reference pixels driving the output.

The honest caveat

Captioning models make confident mistakes, and this one is from the frozen 2024 Fooocus era, so it'll describe the obvious and miss the subtle - the pose, the expression, the thing you actually liked about the image. Treat its output as a strong starting draft, not the answer. And it's a beginner-friendly convenience, not a magic "clone this image" button: for actually transferring style or identity from a reference, the ImagePrompt node is the one that does the real work.