FLUX

FLUX is Black Forest Labs’ open-weight image generation family. The FLUX.1 members are 12 billion parameter rectified flow transformers that turn text descriptions into images. FLUX.2 later added a 32 billion parameter dev model and the smaller klein variants.

The weights live on Hugging Face, and the official black-forest-labs/flux repository ships minimal inference code to run them from the command line or a demo; the pro tier is API only.

The family and its variants

The open-weight catalogue differs from variant to variant, and the licence applies to a specific model, not to the family.

• FLUX.1 [schnell]: 12 billion parameters, distilled for guidance and steps, usable images in 1 to 4 steps.
• FLUX.1 [dev]: also 12 billion parameters, refined with guidance distillation.
• FLUX.1 [pro]: the base model, API only, outside the open weights.
• FLUX.1 Tools [dev]: Fill for in- and out-painting, Canny and Depth for structural conditioning, Redux for variation, Kontext for editing, Krea for text-to-image, plus LoRA builds of Canny and Depth.
• FLUX.2 [dev]: a 32 billion parameter flow matching transformer for generation and reference-based editing; FLUX.2 [klein] brings the fastest variants at 4B and 9B.
• The autoencoder is apache-2.0 on both lines.

Licence: different terms per variant

Commercial use depends on the licence of the specific variant, and the open weights carry usage restrictions.

• FLUX.1 [schnell] is apache-2.0; the model card states it covers personal, scientific and commercial purposes.
• FLUX.1 [dev], the dev tools (Fill, Canny, Depth, Redux, Kontext, Krea) and the Canny and Depth LoRAs fall under the FLUX.1 [dev] Non-Commercial License v1.1.1, which ties non-commercial use to the absence of payment, direct or indirect.
• klein 4B and klein 4B Base are apache-2.0, while the 9B variants and FLUX.2 [dev] fall under the FLUX Non-Commercial License.
• The dev licence claims no rights in the output, but prohibits training, fine-tuning or distilling a competitive model on it and requires content filtering.
• Commercial use of the open weights is licensed separately and reported through the BFL API with the track_usage flag and a BFL_API_KEY variable.

Hardware and running it locally

The official documentation publishes no single VRAM figure; it documents ways of reducing the footprint.

• In the FLUX.1 repository the demos can offload to CPU, and TensorRT exports exist in BF16, FP8 and FP4 precision.
• FLUX.2 [dev] at full precision needs an H100 class GPU. The official diffusers guide quotes roughly 18 GB with 4-bit quantization and a remote text encoder, around 20 GB with a 4-bit text encoder, and above 80 GB everything fits at once.
• FLUX.2 [klein] 4B fits in about 8 GB of VRAM, which the repository places on cards like the RTX 3090 or 4070.
• The FLUX.1 code runs on Python 3.10; the authors tested FLUX.2 on CUDA 12.9 and Python 3.12.

What to watch in practice

• A LoRA adapter counts as a derivative work, so it inherits the licence of its base model, non-commercial terms included.
• The model cards note that the model gives no factual information and that prompt following depends heavily on the prompt style.
• Generative work shares the GPU memory of one machine: the same card serves the video runs of FramePack and the language models in LM Studio, so FLUX settings are fitted to the memory budget.

Further reading

• Black Forest Labs: FLUX repository
• FLUX.1 [schnell] model card
• Commercial licensing

At CyberElectro FLUX is the generative image layer: the fast variant drives iteration, and we check the licence before any client work.