Black Forest Labs, a startup founded by former Stability AI employees, has released Flux, a new family of generative AI models for creating images from text prompts that already delivers capabilities Stable Diffusion 3 was supposed to provide. The release has generated significant attention in the generative AI community, particularly because the team behind it helped develop the latent diffusion technology that powers Stable Diffusion before departing to launch their own venture.

What Makes Flux Different From Stable Diffusion 3?

Flux represents a meaningful leap forward in image generation technology. The model features a new 12-billion-parameter text decoder that allows it to understand and follow text prompts more reliably and accurately than previous generations. Unlike Stable Diffusion 3 Medium, which was trained under strict censorship restrictions that made it nearly unusable for many applications, Flux was developed without the same limitations. This difference is visible in the results: human poses generated by Flux appear more natural, and the model demonstrates improved hand and finger rendering, though this remains an ongoing challenge across the industry.

Another key differentiator is Flux's support for rendering text within images, a capability that Stable Diffusion 3 marketed as a key feature but struggled to deliver effectively. The model can add captions to images exactly as requested, opening new possibilities for designers and content creators who need text integration in their visual outputs.

How Many Versions of Flux Are Available?

Black Forest Labs has released Flux in three distinct versions, each designed for different use cases and deployment scenarios:

• FLUX.1 [pro]: The full version of the model, available exclusively through an application programming interface (API), which is a standardized way for software to communicate. This version requires a subscription and is designed for users who want the highest quality output without managing their own computing infrastructure.
• FLUX.1 [dev]: The largest open-weight model with 12 billion parameters, available for local deployment on personal computers or servers. While the model weights are publicly available for download and further customization, the license restricts non-commercial use; commercial applications require separate negotiation with Black Forest Labs.
• FLUX.1 [schnell]: A distilled turbo version designed for speed, capable of generating images in just four steps instead of the 20 or more steps required by the full version. This model is distributed under the Apache 2.0 license, a permissive open-source license, though technical limitations likely prevent training derivative models from it.

The availability of open-weight versions, particularly the dev and schnell models, represents a significant departure from the API-only approach that many competitors have adopted. This openness appeals to researchers, hobbyists, and developers who want to experiment with the technology without relying on cloud services.

How to Set Up Flux on Your Local Computer?

For users interested in running Flux locally, the setup process requires specific software and hardware considerations. Currently, neither AUTOMATIC1111 nor WebUI Forge, two popular interfaces for Stable Diffusion, support Flux. Instead, the only compatible software at launch is ComfyUI, which can be used either as a standalone application or with the SwarmUI frontend, a more user-friendly interface created by another former Stability AI employee.

The installation process involves several steps:

• System Requirements: Windows users need to install the WinGet package manager from the Microsoft Store and the.NET 8 Software Development Kit (SDK) from Microsoft's website before proceeding with the main installation.
• Model Download: Users must choose between 16-bit or 8-bit versions of Flux depending on their graphics card memory. The 16-bit versions require at least 16 gigabytes of VRAM (video random-access memory), while 8-bit versions work with 12 gigabytes or less, with only slightly lower quality.
• File Organization: Downloaded model files must be placed in specific folders within the SwarmUI directory structure, including the unet folder for the main model and the VAE folder for the supporting encoder file, ensuring the software can locate and use them correctly.

Once these components are installed and configured, users can launch SwarmUI and begin generating images with Flux, though the initial startup requires time for the script to install remaining dependencies.

Why Is Flux Generating So Much Excitement?

The practical results from Flux demonstrate why the model has captured community attention. Early users report that the model handles complex scenarios that previously required extensive trial and error or specialized fine-tuning files called LoRAs. Street scenes render with anatomically correct hands on the first attempt, historical Hollywood-style shots produce five-fingered hands consistently, and children's book illustrations emerge without requiring multiple generation attempts or specialized model variations.

This represents a significant quality-of-life improvement for users. Previous models like SDXL required users to either accept imperfect outputs or spend considerable time searching for and downloading specialized LoRAs, which are customized model variations trained on specific styles or subjects. Flux's broader understanding of concepts and improved rendering capabilities reduce this friction substantially.

The timing is particularly noteworthy because Stable Diffusion 3, which was supposed to represent the next major evolution of the technology, remains in limbo. Flux has effectively filled that gap, delivering the capabilities users expected from Stable Diffusion 3 while avoiding the censorship restrictions that made SD3 Medium problematic for many applications. For the generative AI community, this represents both a technological milestone and a significant shift in the competitive landscape, with engineers who helped build Stable Diffusion now leading the charge with a superior alternative.