FLUX: Black Forest Labs image models, generation by generation

What FLUX is

FLUX is the image generation line from Black Forest Labs. Every open FLUX checkpoint is a rectified flow transformer rather than the U-Net used by earlier Stable Diffusion releases. The line has two generations so far, and each ships both open-weight and API-only variants. This page covers the open-weight ones you can download and run yourself.

Generations in order

FLUX.1 (August 2024)

FLUX.1 was announced on August 1, 2024 with three variants: pro (API only), dev and schnell. Black Forest Labs describes all three as a hybrid architecture scaled to 12 billion parameters, using multimodal and parallel diffusion transformer blocks, flow matching, rotary positional embeddings and parallel attention layers.

What differs between the open variants is mostly how they were distilled and licensed:

• FLUX.1 [dev] is guidance-distilled and is described by its model card as second only to the pro model in output quality. It uses the FLUX.1 [dev] Non-Commercial License. The card says generated outputs can be used for personal, scientific and commercial purposes under the terms of that agreement, but the weights themselves are not freely licensed for commercial deployment.
• FLUX.1 [schnell] is trained with latent adversarial diffusion distillation and generates images in 1 to 4 steps. It is Apache 2.0, so it is the permissive choice in this generation.

Two derivative checkpoints are also 12B and share the dev license. FLUX.1 Kontext [dev] is an image editing model: you give it an image and an instruction, and it supports successive edits with limited drift between them. FLUX.1 Krea [dev] is a text-to-image model whose card says it focuses on aesthetic photography.

FLUX.2 (November 2025)

FLUX.2 was released on November 25, 2025. The flagship open model, FLUX.2 [dev], has 32 billion parameters and pairs the rectified flow transformer with the Mistral-3 24B vision-language model as its text encoder. One checkpoint handles text-to-image and editing, including editing from multiple reference images (the announcement says up to 10) and image editing up to 4 megapixels. Its license is the FLUX Non-Commercial License.

The announcement listed the klein family as coming soon under Apache 2.0. The released klein cards read differently by size:

• FLUX.2 [klein] 4B is Apache 2.0, distilled for few-step generation (the card shows examples with 4 steps), and handles text-to-image and multi-reference editing in one model.
• FLUX.2 [klein] 9B uses an 8B Qwen3 text embedder, is distilled to 4 inference steps, and carries the FLUX Non-Commercial License rather than Apache 2.0.

Check the license on the exact checkpoint before you build on it, because the two klein sizes do not match.

Which one to use today

• Commercial product, permissive license, small footprint: FLUX.2 klein 4B or FLUX.1 schnell. Both are Apache 2.0. The 4B model is the newer design and does editing as well as generation; schnell is a 12B model with a one-to-four step sampler.
• Highest-quality open weights, non-commercial or research use: FLUX.2 dev at 32B is the largest open FLUX. Its size means it needs a much larger memory budget than the 12B models, which the model page computes for your hardware.
• Editing an existing image: FLUX.2 handles editing natively. FLUX.1 Kontext dev is the 12B option if you stay on the first generation.
• Photographic look with the FLUX.1 stack and existing tooling: FLUX.1 dev or Krea dev, subject to the non-commercial license.
• Low latency: the klein models are described as sub-second generators on suitable hardware, and schnell is the few-step option in FLUX.1.

If your pipeline already uses LoRAs or adapters trained for FLUX.1, stay on FLUX.1 until you have checked they load on the new architecture; FLUX.2 uses a different text encoder and a larger transformer.

Side lines and community variants

Black Forest Labs lists no separate side lines in this series beyond the dev, schnell, Kontext, Krea and klein variants above. Community fine-tunes and distilled adapters of FLUX.1 and FLUX.2 are indexed under the generation hubs, FLUX.1 and FLUX.2, next to the first-party checkpoints.

Running FLUX

All of these checkpoints are available in Diffusers and ComfyUI according to their model cards. You can run any of them on Aquanode GPUs.