Structured Output

Prompting an LLM to "return a number" and getting a paragraph back is a rite of passage. This node exists so you stop playing that game. StructuredOutput pins the model to a single, typed value - an int, a float, a bool, a category you define - and guarantees it comes back as exactly that. It's the difference between "the model probably did what I asked" and "the model has no choice but to comply."

What it is

It's the pack's answer to getting structured data out of a GGUF LLM reliably, and it's one of the features the README leads with. The author's framing is blunt: structured outputs are hard to get through prompt engineering alone, so this node does it at the mechanism level. Use it to extract entities, pull numbers, classify prompts against a list of classes you define, or emit one specific prompt - the examples in the README, verbatim.

The inputs that matter

• prompt - the input text you're analyzing: a caption, an image description, whatever you want to extract from.
• model - CUSTOM, the GGUF LLM from the pack's loader.
• attribute_name - the name of the value you want back, e.g. age or style. The model returns it labeled with this.
• attribute_type - the big one. A dropdown: str, int, float, bool, or Category. This is what forces the response into the shape you want.
• attribute_description - a hint describing the field, so the model knows what it's extracting. Filling this well is the difference between accurate extractions and confident guesses.
• categories - comma-separated values, used only when attribute_type is Category. This is your classification target list - e.g. anime, photo, painting.
• temperature - default 0.15. Keep it low for extraction and classification; you want the same answer every time.

Output is a single STRING, but it's the string of the typed value - the clean int, float, or category you asked for, not a JSON blob.

How it works

Under the hood it runs a structured chat: the pack builds a JSON schema from your attribute settings and instructs the model to return only that object. The generation is constrained to follow the schema - that's why you get a number back instead of an essay. Because the schema is built from your attribute_name and attribute_type, the node is genuinely general: flip attribute_type to Category, fill in categories, and the same setup becomes a classifier.

Install

Pack-level install:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI_VLM_nodes

Restart ComfyUI (or Manager → "VLM Nodes"). Python 3.9+ and the llama.cpp stack. You need a GGUF in models/LLavacheckpoints; LLaVa 1.5/1.6 are the README's recommended models for the pack's generation family, so if the model has a vision projector, pair it correctly.

Troubleshooting

Beyond the pack's usual llama-cpp-python install pain, the interesting failure here is semantic: if your extraction is wrong, it's almost never the mechanism - it's your attribute_description being vague, or the prompt being ambiguous. Tighten those first. And if Category returns junk, your categories list is probably too loose; give the model fewer, sharper options.