Prompt Builder#
The Prompt Builder writes a structured JSON prompt for you: subjects, setting, camera, lighting and palette as named fields instead of one long sentence. It is designed for the Flux 2 Klein and Krea 2 families, which read the fields directly; classic SD models see the JSON as text and gain nothing.
Open it from the button next to the prompt box on the Generate tab; it appears as a third panel beside the settings. The builder does not generate by itself. It writes into the prompt box, and the normal Generate button does the rest.
Fields#
| Section | Fields |
|---|---|
| Main prompt | The scene in a sentence: subject, setting, action, details. |
| Subjects | One block per person or object: description, clothing, expression, action, position. Separate subjects keep traits from bleeding into each other in multi-person scenes. |
| Scene | Environment, weather, era, scene action, props, text in image. |
| Look | Style, camera (angle, shot distance, lens in mm, aperture, film stock), lighting, colour palette, mood, composition. |
| Custom fields | Any extra key and value; keys use letters, digits, _ or - and cannot shadow a built-in field. |
Most look fields offer a catalogue of values (lighting setups, film stocks, compositions) and also accept free text. The palette takes colours you add by hand or extracts them from an image.
Presets#
Built-in presets fill the look fields for common photographic and illustrated styles. Save as preset keeps your own field set under My Presets; Update overwrites it with the current fields.
From the builder to the image#
- JSON: the live output; Copy puts it on the clipboard.
- Apply to Prompt writes the JSON into the prompt box. Apply & Generate does that and queues a generation.
- Opening the builder while a JSON prompt is already in the box loads it back into the fields; a plain-text prompt lands in the main prompt field.
- Fill from Image (AI) analyses an image with a vision model and fills the fields from it. The popover picks the model: the vision bundle served by the assistant's LLM server (Models > Assistant, Image to prompt - vision model, 8.8 GB, loaded on demand) or a vision LLM installed on the Image tab (the catalogue offers the Qwen3-VL 4B ones, which fit a 16 GB card; a larger checkpoint you install yourself is refused with a message when it exceeds the card's VRAM). An optional instruction shapes the analysis: a built-in template, one of your saved templates or a custom text, sent to the model as the system message. In the builder a template adds rules; the reply keeps the builder's field schema, so the fields still fill.
Variations#
Pick one field, list two or more values, and Enqueue generates one image per value with a shared seed, so only that field changes between the results. Use it to compare lenses, lighting setups or moods on the same composition.
Rules worth knowing#
- A JSON prompt is sent as-is: dynamic prompt braces are not expanded inside it, and hex colours in the palette are safe.
- Keep the main prompt short and put specifics in the fields; the models weight named fields more reliably than a long sentence.
- The builder is available on the Generate tab only. The canvas uses the same prompt box, so a JSON prompt applied there carries over.