A 5.6B image transformer with native 4K
FLUX 3 Image Klein is Black Forest Labs' open weights image-generation and editing model, released on October 15, 2026. It accepts a text description and optional reference images. You can use it in a hosted demo, call a hosted API, or run a compatible checkpoint on your own hardware.
The published 5.6B count describes the image transformer. A complete installation also needs a text encoder, a decoder and working memory, so checkpoint size alone does not tell you the GPU requirements. “Open weights” also does not mean every use is unrestricted: the model has its own license.
What does native 4K mean?
The model can generate at a 4K output setting rather than requiring a separate upscaling step for every image. The official demo describes 1K as approximately 1 megapixel, 2K as 4 megapixels and 4K as 16 megapixels. The image's aspect ratio determines its exact dimensions. Begin at 1K while refining a prompt; increase resolution after the composition works.
Generate, edit and combine references
Text-to-image
Describe what the image should show, how it should be composed and what kind of finish you want. Keep the official demo's prompt upsampler enabled for ordinary short requests: the publisher expects a full scene caption. Concrete nouns and visual details are easier to judge than a long list of quality adjectives. Treat your first output as a draft and change one important instruction at a time.
Multi-reference editing
Use one photograph as the starting point for an edit, or supply up to 10 images with different roles. For example, request the subject from the first reference, the setting from the second and the palette from the third. In the official demo, the first reference determines the output's shape.
Bounding-box control
The release also includes layout and editing controls based on boxes. These describe where an element belongs or which region should change. The official demo has a separate box workflow; start with its examples instead of assuming ordinary prose precisely fixes every position.
Open the FLUX 3 Image Klein browser demo. Avoid uploading sensitive photographs to a public hosted service; review its privacy terms first.
Base or Distilled: which checkpoint?
| Checkpoint | Starting point | Official demo defaults |
|---|---|---|
| Distilled | Interactive generation and everyday experiments | 16 steps, without classifier-free guidance |
| Base | Flexible sampling and fine-tuning | 50 steps, guidance 3.5 |
These are the current demo defaults, not a promise of equal speed on every machine. Compare quality and time at the same resolution before increasing the number of steps.
Choose how to use the model
Use the browser demo for a first look, or choose a supported local runtime when you want to manage the workflow yourself. A hosted API is a separate service with its own access and billing. Local hardware requirements depend on the runtime, precision and workload.
A visual workflow in ComfyUI
The ComfyUI workflow guide explains native-template setup, component folders and loader troubleshooting. Use it when you want to build and save a visual graph.
Find the right model package
The weights download guide covers repository selection, gated access and file verification. For the full installation order and a first saved output, use how to run the model locally.
License & commercial use
The FLUX Kommunity License governs the released weights. Its qualifying-user definition uses less than US$5 million in annualized revenue, including affiliates, and commercial use carries additional conditions. Read the exact license for your organization and use case; do not treat the threshold alone as complete permission.
The weights are not under the Apache 2.0 terms used by FLUX.2 Klein 4B. A hosted API also has service terms separate from the weights license. For the current binding terms, use the official FLUX Kommunity License and BFL licensing information.
API access and pricing
The BFL endpoint is POST https://api.bfl.ai/v1/flux-3-image-klein. The official quickstart shows text-to-image, editing, multiple references and box-based workflows. An API request returns a task to poll for its result.
Keep API credentials on your server and apply your own usage limits. Do not paste an API key into a public website or client-side script. The native studio keeps its own keys on the server and does not ask visitors to supply one.
The browser demo's free GPU quota is separate from commercial API billing. Check current API pricing and the official request and result contract before integrating.
A practical prompting starting point
For a new image
Write the subject, setting, composition, lighting and finish in that order. For example: “A single ceramic cup on warm limestone, soft morning light from the left, subtle glaze texture, understated editorial photography, no text.” Compare the result with your intended subject before adding more instructions.
For an edit
Name the target and the change, then say what should remain. For example: “In the first image, change only the jacket to deep forest green. Keep the pose, face, background and lighting unchanged.” Inspect the preserved areas as well as the edited object.
For multiple references
Give each reference one role and describe it by upload order. Avoid conflicting instructions such as preserving two different compositions. Start with two references, check the result, then add more only when each contributes something useful.