FLUX 3 Text to Image

FLUX 3 Text to Image (Text-to-Image Model)

What is FLUX 3 Text to Image?

FLUX 3 Text to Image is the text-to-image path of Black Forest Labs' FLUX 3 Image model, the image component of the multimodal FLUX 3 family. From a prompt alone it generates detailed, well-composed stills at resolution tiers from 768sq up to native 4K. Its headline difference from an ordinary text-to-image generator is explicit composition control: elements tagged in the prompt can be pinned to exact regions on a 0-1000 bounding-box grid, so a layout is reproduced rather than inferred from prose. It is built for creators, designers, and developers who need placement, legible text, and print-ready detail from a single API call.

Key Features

• •Bounding-box layout control — tag elements in the prompt and pass a JSON array of ids, boxes, and descriptions on a 0-1000 grid to place each one precisely.
• •Native resolution up to 4K — tiers 768sq, 1k, 1.5k, 2k, and 4k render at the requested size with no separate upscale pass, so fine textures and small text stay sharp.
• •Accurate multilingual text rendering — legible, correctly formed words inside the scene, including non-Latin scripts.
• •Web grounding — an optional toggle lets the model consult web and image search so real places and products render faithfully.
• •Wide aspect-ratio and style range — from 21:9 to 9:21, across illustration, product, painterly, and photoreal looks.

Best Use Cases

FLUX 3 Text to Image is strongest where composition and legibility matter: posters, packaging, and infographics that need text in the right place; product and editorial imagery that must hold detail at print resolution; and concept work where several named elements have to sit in specific positions. In testing, a single prompt placed five distinct components into their exact quadrants with a legible label, and a 4K render kept individual embroidery threads crisp. With grounding on, a prompt naming a real library rendered that interior true to reference.

Prompt Tips and Output Quality

Describe the subject first, then placement. For exact layouts, tag elements with ids and append the bounding-box JSON, keeping to four to six boxes so each has room. Put in-image text in quotes and name its script and style. Choose 4k when the result will be cropped or printed; turn grounding on for real-world subjects and off for pure imagination.

How to Use FLUX 3 Text to Image

FLUX 3 Text to Image turns a written prompt into a single still at resolution tiers from 768sq to native 4K. The prompt does most of the work: describe the subject, then its placement, and the model composes the frame. Two features set it apart — bounding-box layout control and native-resolution rendering — so plan your prompt around them.

Recommended Settings

• •resolution: 1k for quick drafts, 2k for a crisp default, 4k when the image will be cropped or printed. 4K renders at native size with no upscale pass, so fine texture and small text stay sharp.
• •aspect_ratio: auto lets the prompt decide; set a fixed ratio (21:9 to 9:21) for a known canvas.
• •grounding: true for real places, products, or people so they render faithfully; false for pure imagination.
• •safety_tolerance: leave at 2.
• •Layout: tag elements with <id> in the prose and append a JSON array of {id, bbox:[y_min,x_min,y_max,x_max], desc} on a 0-1000 grid. Keep to four to six boxes so each element has room.

Example prompts to try

Each example below ran end to end at the settings shown; the layout, resolution, text, and real-place subject each carried over into the output.

• •Bounding-box layout (default). Prompt: "Top-down flat-lay on a reclaimed-wood workbench, four components arranged in a clean grid with even daylight: a disassembled brass pocket-watch movement <watch>, a tight coil of bright copper wire <wire>, a teal printed circuit board <board>, and a fan of red-handled jeweler's screwdrivers <drivers>, with a small white index card <label> in the centre." — followed by the bbox JSON array pinning each tagged id. All five elements landed in their quadrants and the card read "BENCH KIT No. 3". Tip: keep to four to six boxes so each element gets room.
• •Native 4K fine detail. Prompt: "Extreme close-up of a hand-embroidered silk panel, gold metallic thread forming a peacock-feather motif over deep indigo fabric, a steel needle and a loose thread resting on the woven surface, even daylight." At 4k each thread and the weave stayed crisp. Tip: use 4k when the result will be cropped or printed.
• •Accurate multilingual in-image text. Prompt: "A weathered hand-painted wooden shop sign above a spice merchant's stall in an old bazaar. The sign reads 'मसाला भंडार' in large ochre Devanagari letters, with 'SPICE HOUSE - EST. 1911' beneath in cream serif capitals. Burlap sacks of turmeric and dried red chillies sit below, even daylight." Both scripts rendered legibly. Tip: quote the exact words and name the script and placement.
• •Web grounding (real named place). Prompt: "The barrel-vaulted iron-and-glass reading room of the Bibliotheque Sainte-Genevieve in Paris, long rows of wooden desks under green-shaded brass lamps, tall arched windows along the walls, late-afternoon light." With grounding: true the real reading room rendered true to reference. Tip: name the actual place and leave grounding on.

Further prompts the model is built for: "A product packaging label with the brand name placed in the top third", "A travel poster with a city name across the bottom", "A storefront sign in two languages", "A labeled cutaway diagram with parts pinned by bounding boxes".

Can it place elements at exact coordinates?

Yes, using bounding boxes on a 0-1000 grid with ids and descriptions.

Does it produce real 4K?

Yes, rendered natively at the requested tier rather than upscaled afterward.

Can it render text, including other languages?

Yes, including multilingual text; results are best when you quote the exact words.

Can it edit an existing photo?

Not on this text-to-image endpoint; reference-image editing uses the separate FLUX 3 Image edit path.

What does grounding do?

It lets the model consult web and image search so real places, products, and people render faithfully.