FLUX Inference Codebase

FLUX.1 Kontext [dev] is a 12 billion parameter rectified flow transformer designed for instruction-based image editing.

Core Capabilities:

• Image Editing: Modify existing images using text instructions.
• In-Context Reference: Use characters, styles, or objects as references without requiring fine-tuning.
• Iterative Refinement: High consistency allows for multiple successive edits with minimal visual drift.
• Efficiency: Trained using guidance distillation for improved performance.

Implementation Options:

• Reference Implementation: Available in the black-forest-labs/flux repository.
• Integrations: Supported in ComfyUI and Diffusers.

FLUX.1 Krea [dev] is released under the FLUX.1 [dev] Non-Commercial License.

Key Licensing Details:

• Permitted Use: Generated outputs can be used for personal, scientific, and commercial purposes only as described in the specific license terms (refer to the FLUX.1 [dev] Non-Commercial License).
• Prohibited Use: The license prohibits the generation of unlawful content or the use of generated content for unlawful, defamatory, or abusive purposes.
• Compliance: Developers using the open-weight model are expected to implement filters or manual reviews for illegal or infringing content to comply with the license terms.

Licensing

FLUX.1 [dev] is released under the FLUX.1 [dev] Non-Commercial License. While outputs can be used for personal, scientific, and commercial purposes as defined by the license, users must adhere to strict usage policies.

Out-of-Scope Use

The model and its derivatives may not be used for:

• Violating any laws or regulations.
• Exploiting or harming minors.
• Generating verifiably false information to harm others.
• Generating personal identifiable information (PII) to harm individuals.
• Harassment, abuse, threats, stalking, or bullying.
• Creating non-consensual nudity or illegal pornographic content.
• Fully automated decision making that impacts legal rights.
• Facilitating large-scale disinformation campaigns.

FLUX.1 Redux is an adapter designed for the FLUX.1 text-to-image base models (FLUX.1 [dev] and FLUX.1 [schnell]). It allows you to generate image variations based on an input image.

Available Model:

• FLUX.1 Redux [dev]: Available at black-forest-labs/FLUX.1-Redux-dev. Note that this model is subject to the FLUX.1-dev Non-Commercial License.

If you prefer not to host the model locally, FLUX.1 [schnell] and other FLUX models are available through several API providers:

• fal.ai: Specifically offers fal-ai/flux/schnell.
• replicate.com: Provides access via their Flux collection.
• bfl.ml: Currently provides FLUX.1 [pro] via API.

By default, weights are downloaded automatically to the checkpoints/ directory from HuggingFace when running demos. To use manually downloaded weights, you can place them in checkpoints/ or point to them using the following environment variables:

• FLUX_MODEL: Path to your model file.
• FLUX_AE: Path to your autoencoder file.

export FLUX_MODEL=<your model path here>
export FLUX_AE=<your autoencoder path here>

For TensorRT support, you must use an NVIDIA PyTorch image via enroot. Follow these steps to import the image, create a container, and install the repository with the [tensorrt] extra and the NVIDIA PyPI index.

cd $HOME && git clone https://github.com/black-forest-labs/flux
enroot import 'docker://$oauthtoken@nvcr.io#nvidia/pytorch:25.01-py3'
enroot create -n pti2501 nvidia+pytorch+25.01-py3.sqsh
enroot start --rw -m ${PWD}/flux:/workspace/flux -r pti2501
cd flux
pip install -e ".[tensorrt]" --extra-index-url https://pypi.nvidia.com

Streamlit Demo

Provides both text-to-image and image-to-image capabilities.

streamlit run demo_st.py

Gradio Demo

An interactive experience with the following options:

• --name: Model to use (flux-schnell or flux-dev)
• --device: Device to use (default: cuda if available, otherwise cpu)
• --offload: Offload model to CPU when not in use
• --share: Create a public link to your demo

# Run Gradio with dev model and a public link
python demo_gr.py --name flux-dev --share

python demo_gr.py --name flux-schnell --device cuda

FLUX models are compatible with the diffusers library.

1. Install diffusers: pip install git+https://github.com/huggingface/diffusers.git
2. Use FluxPipeline to load and run the model.

Tip: Use pipe.enable_model_cpu_offload() to save VRAM if you have limited GPU power. For FLUX.1 [dev], use a larger number of num_inference_steps than for schnell.

import torch
from diffusers import FluxPipeline

model_id = "black-forest-labs/FLUX.1-schnell"

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload() 

prompt = "A cat holding a sign that says hello world"
seed = 42
image = pipe(
    prompt,
    output_type="pil",
    num_inference_steps=4, 
    generator=torch.Generator("cpu").manual_seed(seed)
).images[0]
image.save("flux-schnell.png")

When using open-weight models commercially, you must track usage via the BFL API. To do this, set your BFL_API_KEY environment variable and use the --track_usage flag in the CLI. This applies to FLUX.1 Kontext [dev], FLUX.1 [dev], and FLUX.1 Tools [dev].

export BFL_API_KEY="your_api_key_here"

# For a single generation with tracking
python -m flux kontext --track_usage --prompt "replace the logo with the text 'Black Forest Labs'"

# For continuous loop with tracking
python -m flux kontext --track_usage --loop