JSON Parser (FeiMao-326)

The annoying thing about asking an LLM for structured output is that you get a JSON blob, and ComfyUI doesn't care about JSON blobs - it cares about wires. You want the subject on one wire, the style on another, the scene on a third. That's exactly what this node is for: it takes a JSON payload, digs through it, and routes each key you ask for to its own output.

How it works

Give it json_payload (any JSON string) and keys_to_extract (a comma-separated list, like shot, subject, scene, audio), and it does a recursive search through the whole structure. The search is deep - it doesn't just look at the top level. If your payload nests subject three layers down inside some array, it still finds it. Every value it finds under a key gets joined with ", ", so a key containing an object or a list becomes a tidy comma-separated string.

That depth-first behavior is the thing to love about it. LLM output formats drift run to run; a parser that demands an exact schema breaks the moment the model reorders a key. This one shrugs and finds the data anyway.

The node exposes up to 50 outputs (output_1 through output_50), but you don't have to deal with all of them - a bundled JS extension shows and hides outputs based on which ones you actually wire up. So the practical view is "one output per key you asked for."

Wiring it into a real workflow

This node was clearly designed to sit right after an LLM call, and the pack's own recommended pipeline is:

General API Node (force_json_format ✅) → Markdown Extractor → JSON Parser

With force_json_format on, the API returns clean JSON, Markdown Extractor strips any residual code fences, and then you parse. The default payload in the node is a nice demo of the intended use: a shot list for a video, with shot, subject, scene, and audio each holding their own little nested dict. Wire output_1 (shot) into a prompt-template's subject slot, output_2 into the style slot, and you've got a workflow where the LLM literally writes your prompts piece by piece.

Inputs and outputs

Only two inputs, both strings: json_payload and keys_to_extract. That's it. No tuning, no regex, no schema definition - which is either refreshing or limited depending on your mood. If your key has a space in it or you need nested paths like metadata.camera, you're out of luck; it matches plain key names.

The outputs are output_1 … output_50, each a STRING. Keys you ask for but don't find produce empty strings rather than errors, which is graceful but means you should sanity-check a wire that comes back empty - it might be a missing key, not a blank value.

Gotchas

Invalid JSON doesn't crash your workflow. It returns ERROR: Invalid JSON format. ... in output_1 and empty strings everywhere else, so an LLM that decides to be creative with braces will show up as a very visible error string rather than a silent failure. Wire that first output into a Show Text node and you'll see it instantly.

Installing

Same pack as the rest: it's FeiMao_326_JsonParser in the FeiMao-326 category. ComfyUI Manager (search "Comfyui-General-API-Node") or:

cd ComfyUI/custom_nodes
git clone https://github.com/FeiMao-326/Comfyui-General-API-Node.git
cd Comfyui-General-API-Node
pip install -r requirements.txt

Then restart ComfyUI. Dependencies are just openai, numpy, Pillow, requests - nothing heavy, no models to download.