🪠️ WWAA JSON Prompt Builder

Prompt syntax has moved on. Modern LLM-encoded models like Qwen-class and Flux.2 read natural language, and the current advice for structured, repeatable generation is to keep your fields separate - which is exactly what JSON does. WWAA JSON Prompt Builder is a giant form that turns ~30 labelled fields and dropdowns into one clean JSON string, so you get consistent, schema-shaped prompts without hand-typing braces.

The mental model is a photography brief: you fill in the scene, the subject, the environment, the lighting, and the output parameters, and the node assembles it all. Its defaults are suspiciously specific - a retro balloon photo shoot with direct flash and a disposable camera look - which tells you this was built for a particular character-photo workflow at WeirdWonderfulAI.Art. Ignore those defaults and it becomes a general structured-prompt generator.

The inputs that matter

There are a lot, but you'll realistically touch these:

• scene_description - the free-text scene, the one required field you type into.
• subject_category - human/animal/object/abstract/landscape/architecture; it decides whether the subject-detail fields get used.
• enable_subject_details - master switch for the whole subject block (hair, face, pose, clothing…). Flip it off for pure landscapes and the form stops shoving empty fields at you.
• lighting, mood, photography_style, camera_angle - dropdowns with sensible presets, each paired with a *_custom text box for when a preset doesn't fit.
• aspect_ratio and render_intent - output-level metadata, useful if you're piping this JSON to an LLM or a workflow that parses it.
• custom_fields - free-form extra key-value pairs (newline separated) when the form genuinely lacks a field.
• include_empty_fields and indent_json - the first drops empty strings from the output to keep it lean; the second controls pretty-printing. If you're feeding a captioner or LLM that's picky, indent_json: false gives you compact single-line JSON.

Output is a single json_prompt string, which you'd usually wire into a prompt encoder or an LLM node that understands structured input.

Install

In ComfyUI Manager, search "WWAA Custom Nodes" and install. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/hgabha/WWAA-CustomNodes
# restart ComfyUI

The pack declares zero extra requirements - numpy, Pillow, torch only, all stock - so there's no dependency roulette. Look under the 🪠️ WWAA menu.

Should you use it?

Honest take: if you write prompts by hand and like it, this form will feel like bureaucracy. It earns its keep in two scenarios - when you want consistent, repeatable structure across hundreds of generations (LoRA dataset captions, batch photography-style work), and when a downstream LLM or captioner genuinely consumes JSON and you want field separation you can rely on. For everyone else, it's a fun toy and a decent way to see how a professional studio structures a prompt. Just remember: the JSON is a prompt, not magic. A good JSON brief still needs good words inside it.