JSON to Text

The prompt-generation nodes in this pack return JSON-looking text, and JSON is not what you want sitting in a CLIP encoder. JsonToText is the small utility that bridges that gap: it takes that string, parses it, and hands you back the same content as readable text. It's not glamorous, and it's not supposed to be.

What it is

JsonToText lives under the pack's Text category and is a display/output node. You drop it after any node that emits structured text - LLavaPromptGenerator, StructuredOutput, Suggester, CreativeArtPromptGenerator - and it renders the JSON as human-readable text so you can see what the model actually decided, or feed that text onward.

The one input

• text - a STRING holding JSON, possibly wrapped in code fences, possibly buried inside a larger response. That's the entire input surface. No options to tune, no hidden settings that matter for day-to-day use.

The output is a single STRING. Since it's an output node, you'll typically wire that string into a text display or use it as the final stop in your prompt-engineering branch.

How it works

It parses plain or fenced JSON and renders it as readable text - values pulled out, key/value pairs laid out sensibly, nested JSON pretty-printed so you can actually scan it. If the input isn't valid JSON, it doesn't blow up; it passes the text through so you always get a string back. That "never fail on weird input" behavior is the whole point of a utility like this in a chain where LLM output is never guaranteed to be clean.

Install

Same as everything in this pack:

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

Restart ComfyUI and it's there. ComfyUI Manager users can just search "VLM Nodes." Python 3.9+.

When you'd skip it

If you're only using the plain LLaVa samplers or LLMSampler, you don't need this node - those return plain strings already. Reach for JsonToText specifically when something upstream hands you a JSON-looking blob. That's its entire job, and it does it without fuss.