FLUX.1 - BlackForestLabs

BlackForestLabs is proud to announce the release of the FLUX.1 suite, a groundbreaking set of text-to-image models that set a new standard in the field. The FLUX.1 suite excels in image detail, prompt adherence, style diversity, and scene complexity, establishing a new state-of-the-art for text-to-image synthesis.

Designed with accessibility and versatility in mind, FLUX.1 comes in three distinct variants: FLUX.1 [pro], FLUX.1 [dev], and FLUX.1 [schnell].

Try FLUX.1

FLUX.1 Variants: Tailored for Different Needs

FLUX.1 [pro]
[pro] represents the pinnacle of text-to-image synthesis, delivering unmatched performance in image generation. It offers superior prompt adherence, exceptional visual quality, intricate image detail, and diverse output. This variant is designed for users who demand the best. FLUX.1 [pro] is available via API, as well as on platforms like Replicate and fal.ai. For enterprises seeking customized solutions, dedicated support is also available.
FLUX.1 [dev]
[dev] is an open-weight, guidance-distilled model tailored for non-commercial use. Distilled directly from FLUX.1 [pro], it maintains similar quality and prompt adherence while offering greater efficiency. FLUX.1 [dev] is available on Hugging Face, and users can experiment with it on Replicate or Fal.ai. This model provides a powerful option for developers and researchers working in non-commercial environments.
FLUX.1 [schnell]
[schnell] is the fastest variant in the suite, optimized for local development and personal use. Released under the Apache 2.0 license, it provides an accessible option for individuals and small teams. Weights for FLUX.1 [schnell] are available on Hugging Face, with inference code accessible on GitHub and Hugging Face’s Diffusers. Additionally, FLUX.1 [schnell] boasts day-one integration with ComfyUI, making it a convenient choice for developers.

Advanced Architecture: Pushing the Limits of AI

All FLUX 1 models are built on a hybrid architecture that combines multimodal and parallel diffusion transformer blocks, scaling to an impressive 12 billion parameters. By leveraging flow matching, a simple yet powerful method for training generative models, FLUX 1 surpasses previous state-of-the-art diffusion models. The suite also incorporates rotary positional embeddings and parallel attention layers, which significantly enhance model performance and hardware efficiency. A detailed technical report on these innovations will be published soon.

Setting a New Benchmark in Image Synthesis

FLUX.1 establishes a new benchmark in the field of image synthesis. The [pro] and [dev] variants outperform popular models like Midjourney v6.0, DALL·E 3 (HD), and SD3-Ultra in key areas such as visual quality, prompt following, size/aspect variability, typography, and output diversity.

FLUX.1 [schnell] stands out as the most advanced few-step model, surpassing not only its direct competitors but also non-distilled models like Midjourney v6.0 and DALL·E 3 (HD).

Each FLUX.1 model is finely tuned to preserve the full output diversity achieved during pretraining, offering enhanced creative possibilities for users.

The Future of Text-to-Video with FLUX.1

With the release of the FLUX.1 text-to-image suite, BlackForestLabs sets the stage for the next frontier in generative media. These models lay the groundwork for an upcoming suite of competitive text-to-video systems, designed to deliver precise creation and editing at high definition and unprecedented speeds. As BlackForestLabs continues to pioneer the future of generative media, users can expect more innovative tools that push the boundaries of what’s possible in AI-driven creativity.

Stay tuned for more updates as we continue to explore the future of generative AI and media creation.

Try FLUX.1

Inference partners

We are happy to partner with Replicate and FAL. You can sample the models using their services. Below we list relevant links.

Replicate:

FAL: