Architectural Rendering with FLUX 3 Image: Text-to-Image, Editing and Reference Techniques

FLUX 3 Image is the image generation and editing component of Black Forest Labs' FLUX 3 family, offering text-to-image generation, editing existing images, merging multiple reference images, and composition control via bounding boxes—all through a single API endpoint. This guide walks through these four core use cases step by step for those looking to produce architectural presentation visuals and interior design imagery.

What you'll learn

By following this guide, you'll see how to generate an architectural exterior or interior image from scratch using text, how to change only a specific detail (material, lighting, furniture) on an existing render, how to merge a site photo with a reference object, and how to use bounding boxes to lock down the position of every element in a composition.

Requirements

• A Black Forest Labs API key (sent as x-key in the request header)

• Access to the POST https://api.bfl.ai/v1/flux-3-image endpoint; the only required parameter is prompt

• For those who want to use references, 1 to 10 images (URL or base64, between 256×256 pixels and 16 MP)

• aspect_ratio selection (various ratios from 21:9 to 9:21, or auto)

• resolution selection: 768sq, 1k, 1.5k, 2k, or 4k

• Information on safety_tolerance (0 strictest, 4 most permissive, default 2) and grounding (enabled by default) parameters

Step-by-step usage

1. Text-to-image generation

To generate an architectural scene from scratch, it's important to clearly describe the subject, setting, lighting, and framing. FLUX 3 Image builds the entire scene from a single prompt.

Contemporary single-family house exterior at golden hour, board-formed concrete facade, large glazed corner window, cantilevered wooden canopy, landscaped garden in foreground, soft warm sunlight, photorealistic architectural visualization, wide-angle framing

Since this prompt provides material (concrete, glass, wood), lighting condition (sunset), and framing (wide-angle) together, it helps the model construct the scene consistently.

2. Editing an existing image

To change only a specific detail on a render, the instruction should clearly name the element to be changed; anything not specified in the prompt remains unchanged.

Change the facade cladding material from board-formed concrete to vertical timber slats, keep the window layout, lighting, and landscaping unchanged

This approach is suitable for offices that want to test a single variable—such as facade material—while keeping the same composition.

3. Merging multiple reference images

Up to 10 reference images can be uploaded, with each image's purpose specified by referring to its position. For example, a site photo can provide the scene while a second image provides the furniture to be added to the scene.

Use image 1 as the existing living room interior and image 2 as the armchair reference. Place the armchair from image 2 into the corner of the room shown in image 1, matching the room's existing light and perspective

This technique can be used to realistically place a furniture catalog image sent by a client into an existing interior photo.

4. Composition control with bounding boxes

Bounding boxes added at the end of the prompt define the position of each element in the scene using [top, left, bottom, right] coordinates on a 0 to 1000 grid. The same boxes can also be used to recolor, move, or remove only a portion of the scene—such as a single piece of furniture.

Open-plan office interior with a reception desk and a seating area. Reception desk at [100, 50, 400, 500]. Seating area with two armchairs at [450, 550, 850, 950]

Thanks to bounding boxes, a designer can directly control a layout that closely follows a floor plan during the image generation stage.

5. Using the grounding setting

Grounding is enabled by default, and the model searches the web to generate real-world objects more accurately; when disabled (grounding: false), results are produced faster based solely on the prompt text. Grounding should be kept on when accuracy for a specific product or brand matters, and can be turned off for quick concept experiments.

Common mistakes

• Generating without specifying an aspect ratio; the model defaults to 1:1 if there's no reference, which may not suit wide exterior framings.

• Not specifying what should remain unchanged in editing prompts; this can cause the model to alter more elements than intended.

• Writing bounding box coordinates on a scale other than the 0-1000 grid, such as in pixels.

• Trying to download the result before the status field on polling_url reaches Ready, or trying to retrieve the result after the 1-hour download window has expired.

• Stacking many reference images (up to 10) into a single prompt without referring to each one; the position of each image must be explicitly addressed.

Next steps

The browser-based Playground can be used to experiment without writing code; those who want to deepen their prompt structure for scene, style, and text generation can refer to the official prompting guide, and those who want to detail bounding box usage can check the bounding box documentation. The API reference should be consulted for all request parameters and error codes.

Source and license

This guide has been adapted into Turkish based on parameter and endpoint information from Black Forest Labs' official API documentation (docs.bfl.ai, FLUX 3 Image Overview) and general capability descriptions from the company's official product pages (bfl.ai); the example prompts have been written originally for architectural and interior use cases.