Flux

Flux is an image generation and editing model family open-sourced in August 2024 by Black Forest Labs (BFL), a Freiburg / Europe-based startup founded by several of the original Stable Diffusion authors (Robin Rombach, Andreas Blattmann, Björn Ommer, and others). Flux departs from the traditional UNet-diffusion design in favor of a Transformer architecture based on flow matching, delivering materially better prompt adherence and text rendering than contemporaneous SDXL-family models. The release ships three layers: the open-source inference repo on GitHub, the model weights on HuggingFace, and BFL's commercial cloud API at docs.bfl.ai.

Core capabilities center on "open weights + a multi-size, multi-task model family". FLUX.1 [schnell] is the 4-step distilled fastest tier, released under Apache-2.0 with commercial use permitted; suitable for real-time previews and high-throughput generation. FLUX.1 [dev] is the 50-step guidance-distilled quality tier, released under the FLUX.1-dev Non-Commercial License (research / non-commercial only); stronger instruction following and quality than schnell. Editing and conditioning family: FLUX.1 Fill [dev] (in/out-painting for localized repaint and canvas extension), FLUX.1 Canny [dev] and FLUX.1 Depth [dev] (structural conditioning via Canny edges or depth maps, plus corresponding LoRA adapters), FLUX.1 Redux [dev] (image variation from a reference), FLUX.1 Kontext [dev] (image editing driven by natural-language instructions), and FLUX.1 Krea [dev] (a Krea-collaborated text-to-image variant). The autoencoder weights are released separately under Apache-2.0. Deployment shapes: standard install via git clone + Python 3.10 venv + pip install -e .[all]; an optional TensorRT path inside the NVIDIA NGC PyTorch 25.01 container (pip install -e .[tensorrt]); commercial usage tracking via the --track_usage flag integrated with BFL's API. Cloud API at docs.bfl.ai exposes higher-tier closed-source models such as FLUX Pro; commercial licensing happens through bfl.ai/pricing/licensing.

Use cases include: open-weight researchers running ablations and exploring new fine-tuning recipes (LoRA / ControlNet / IP-Adapter adapted to Flux), AIGC engineers embedding schnell into self-hosted products (Apache-2.0 commercial use), independent creators who need high prompt adherence and reliable text rendering (Flux materially outperforms SDXL on long prompts and brand-typography rendering), and B2B integrations reaching higher-tier models (FLUX Pro and beyond) through BFL's API. Hardware floor: standard install recommends an NVIDIA GPU with 8GB+ VRAM; CPU works but is very slow; the TensorRT path is best on post-Ampere NVIDIA GPUs. Language: the README is English-only and community docs are primarily English, but the model itself handles Chinese prompts reasonably well thanks to multilingual training data.

Licensing notes: FLUX.1 [schnell] and the autoencoder are Apache-2.0 and commercially usable; *FLUX.1 [dev] and the entire -dev family (Fill / Canny / Depth / Redux / Kontext / Krea) are under the FLUX.1-dev Non-Commercial License and cannot be used commercially without a separate agreement — for commercial use, route through BFL's API or purchase a license at bfl.ai/pricing/licensing. The repository has not received a new push since July 31, 2025 (the most recent commit corresponds to TensorRT path docs), but BFL continues to ship new models and versions on HuggingFace and via the API.