FLUX.1: Stable Diffusion creators launch new image AI

Black Forest Labs unveiled FLUX.1 on Thursday, a family of models that can generate images from written instructions. The company was founded by several researchers who worked on Stable Diffusion, one of the systems that opened AI image creation to developers and individual users.

The launch matters for two reasons: it comes from a team with a decisive track record in the field and offers a large-scale alternative to the closed models from OpenAI, Google and Midjourney. But it also shows that the open-model label no longer means just one thing: FLUX.1 combines a commercial version with others whose weights — the files needed to run the model — can be downloaded under very different terms.

Three versions for different needs

The family consists of FLUX.1 [pro], FLUX.1 [dev] and FLUX.1 [schnell]. All three are built around an architecture trained with 12 billion parameters, an approximate measure of a model’s ability to detect patterns during training.

The pro version is the highest-quality product and is offered through an API — a service that other applications can connect to on a pay-per-use basis. This is the route Black Forest Labs is taking to compete in the professional image-generation market.

FLUX.1 [dev], by contrast, lets users download the weights for research and noncommercial use. That is a significant concession for artists, researchers and small teams that want to test the system on their own machines or servers, although its license does not allow them to turn it directly into a commercial product.

The third option, FLUX.1 [schnell], is designed for speed and distributed under the Apache 2.0 license, a permissive software license that allows commercial use. The trade-off is that it aims to produce images with fewer computational steps than the highest-quality models. For a tool that generates many previews or runs inside an application, that difference in cost and speed may matter more than a perfect finish.

Beyond Stable Diffusion

Since 2022, Stable Diffusion has popularized a powerful idea: an image AI could be installed, modified and run outside a major platform’s infrastructure. That helped spur a community of interfaces, specialized models and editing tools, but it also made clear that opening up model weights does not, by itself, resolve issues such as the provenance of training data, copyright or the generation of misleading content.

Black Forest Labs is now trying to improve one of the most visible limitations of earlier generations: following complex instructions. According to the company’s published evaluations, FLUX.1 outperforms Midjourney v6.0, DALL·E 3 and Stable Diffusion 3 Ultra in visual quality and prompt adherence. These comparisons were produced by the vendor itself, so they will need to be tested against independent evaluations and real-world use.

The company attributes the improvement to a hybrid architecture combining multimodal transformer and diffusion blocks, along with a technique called flow matching. Put simply, the model learns to progressively turn noise into a coherent image and better connect the words in a prompt with the visual elements it needs to create.

That should be apparent in scenes with multiple objects, spatial relationships and text inside the image — three areas where many generators still make obvious mistakes. It does not mean deformed hands, illegible labels or overly literal interpretations of ambiguous prompts will disappear; it means the system starts from a more ambitious foundation for solving them.

A new split in image AI

The launch comes after a difficult few months for Stability AI, the company that commercialized Stable Diffusion and where several members of the new team previously worked. Black Forest Labs is positioning itself as an attempt to rebuild research and product capabilities under a different structure, in a sector where training cutting-edge models requires growing investment in data, chips and servers.

For users, FLUX.1 expands the options: there will be a fast, reusable version for integrations, a development option for experimentation and a professional service for higher-end results. For the industry, the question will be whether Black Forest Labs can sustain that combination of partial openness and an API business without repeating the financial dependence that has pushed other open projects toward increasingly closed models.

The weights for FLUX.1 [dev] and FLUX.1 [schnell] are available on Hugging Face, while the pro version can be used through the Black Forest Labs API and infrastructure partners. The coming weeks will show whether its results hold up outside controlled demonstrations and what tools the community that made Stable Diffusion a de facto standard builds around it.