The many faces of Flux

Flux, from Black Forest Labs, is a cutting-edge generative AI model that comes in 3 versions:

• Flux.1 Pro can only be used on-line. If you want to visit a website and try generating some image with Flux, go to FLUX.1 [pro] | AI Playground (requires log-in to generate, but lets you generate about 20 images for free.) This commercial version of Flux is also available in paid services such as X's Grok AI, and through an API that could be used by other businesses.
• Flux.1-Dev is the highest quality base model that you can download and run locally on your own computer. Of course, you need a Stable Diffusion interface to use it—right now SwarmUI or ComfyUI are the first to support Flux. Using what was traditionally a "Stable Diffusion" interface with Flux is a huge upgrade that a lot of people are excited about, if they have a system capable of running this large a model.
• Flux.1-Schnell is also available for download (and can be run on-line.) 'Schnell' is German for 'Fast', so this is the "Fast Flux," optimized to produce fairly high quality images in only 4 steps instead of 20 or 30 steps.

These versions mean that Flux is a new competitor against on-line models like Midjourney and DALL-E 3, but also serves as a next-generation upgrade from Stable Diffusion, for people who generate images locally on their own computers. The freely downloadable versions are already being modified by members of the open-source community, who are developing LoRAs to give it different looks, ControlNet models for guiding Flux generations, and even more compact models such as the scaled-down FLUX.1-dev-fp8 which is ideal if your graphics card doesn't have enough VRAM to run the other models.

Flux is great at adhering to long, detailed prompts, rivaling DALL-E 3 in interpreting plain-English descriptions of scenes and characters. Flux also works well in a range of aspect ratios, so if you want tall images for social media, wide-screen images, or anything in between, Flux has more choices than DALL-E 3.

If all you want is a beautiful picture, then even very short prompts can work well—you can enter a one-word prompt such as "Bang!" and try your luck. But if you want a specific look, it's better to describe a person or scene in detail. Some people even use a large language model like Chat GPT to write detailed prompts or descriptions, then paste these into Flux.

Flux is quite good at generating text within scenes. If your prompt includes text that is supposed to appear in a sign, on a t-shirt, etc. then Flux can generate text about as reliably as Ideogram, which is also famous for being good at text.

Flux Pro is not heavily censored, although it has not been trained on details of creating convincing nude models (recent user-trained loras are improving Flux's capabilities for this, but they are not built into the Pro version.) When generating people, Flux is not very good at giving you a specific age that's mentioned in the prompt, although adding description of the person's appearance can create people of all ages.

As with Midjourney, Flux is a model trained to produce beautiful results on the first try, instead of requiring lots of time tweaking and retrying. Unlike Midjourney, Flux isn't yet available on-line with a full suite of tools for inpainting, outpainting, creative upscaling, and all the other tools you often need after initially generating an image (although you can already get all of that in interfaces like SwarmUI that support Flux locally.) 

The output of Flux.1 Pro are limited in resolution, outputting images only about 1024x1024 resolution, depending on the aspect ratio. If you are only using Flux through online interfaces, your best bet for taking an image to a higher resolution would be to use another online service like KREA.AI, which contains an excellent up-scaler that adds appropriate detail to scenes.

Flux seems to be developed primarily in pursuit of photorealism. It's great at making realistic images that looks like photographs or scenes from a movie. Flux can also do anime images and the looks of other media and art styles, but it seems to prioritize following other parts of the prompt, and sometimes doesn't adhere to the artistic style described in a prompt. Loras (helper models) are just starting to be released which could help make Flux into more of an expert in different looks and styles, so I look forwards to trying more of those as they come out.

Flux has a Guidance parameter you can adjust which influences the visual style of its output. Values close to 1 (whether they are exactly 1, or something near it like 1.4) give you more stylized, painterly output, but can also reduce contrast and clarity in the scene, and make the scene more chaotic. As you move into higher values, the look becomes more normal and realistic.

While Flux is brand-new at the time of this review, so there aren't as many bells and whistles added to Flux-based interfaces online yet, it is already one of the most powerful models available, and well worth a try, whether you are just using it online, or you want to download it and use it on your own computer.