FLUX Models Compared: schnell, dev and FLUX 2 Pro

You write a careful prompt for a poster: a coffee shop sign that says "OPEN LATE", warm window light, rain on the pavement. The image looks great, except the sign reads "OPN LAIE". You switch models, change nothing else, and the sign is suddenly perfect. Same prompt, different result.

That gap is the main reason to understand the FLUX models instead of treating them as one thing. They share a name and a maker, Black Forest Labs, but they're built for different jobs. Here's how FLUX.1 schnell, FLUX.1 dev and FLUX 2 Pro differ, and how to pick between them.

The short version

• FLUX.1 schnell: the fast one. Good for drafts and testing ideas.
• FLUX.1 dev: the middle one. More considered output than schnell, with a stricter license.
• FLUX 2 Pro: the newer generation. Aimed at finished work, with better text and reference-image support.

Treat them as three stages of one workflow, not three rivals.

FLUX.1 schnell: fast drafts

According to its Hugging Face model card, schnell is a 12 billion parameter model "trained using latent adversarial diffusion distillation," which lets it generate in 1 to 4 steps. The example code uses 4 steps and a guidance scale of 0.0. It's released under Apache 2.0, which permits personal, scientific and commercial use.

In plain terms: it's built to be quick. That makes it the right tool when you don't know what you want yet.

What it's good for:

• Testing five different compositions before committing to one.
• Checking whether a concept reads at all (a product on a table, a character in a doorway).
• Trying wording changes quickly and cheaply.

Where it's weak: the model card itself says prompt adherence varies with phrasing and that outputs may fail to match what you asked for. In my experience that shows up most with long, multi-part prompts and with small text. Treat a schnell result as a sketch, not a final.

FLUX.1 dev: a step up in care

The FLUX.1 dev model card describes a 12 billion parameter rectified flow transformer trained with guidance distillation. The example code uses 50 inference steps and a guidance scale of 3.5. Those are suggested values, not rules.

Two things matter for creators:

1. More steps, slower runs. Where schnell is built for a handful of steps, dev's example code uses far more. Expect it to be the slower of the two when you compare them on the same platform.
2. The license is different. The weights are released under the "FLUX.1 [dev] Non-Commercial License Agreement." If you're making work for clients or selling products, read the license terms yourself before assuming anything. This applies to the downloadable weights; how a hosted service handles output is that service's own policy, so check it.

Dev shares the same caveats as schnell: it can't reliably give factual information, it can amplify biases, and it may miss your prompt. It's a good "second pass" model: once schnell has shown you a direction, rerun the prompt on dev and compare.

FLUX 2 Pro: the finishing model

Black Forest Labs' FLUX.2 announcement positions FLUX.2 [pro] as delivering image quality that "rivals the best closed models," while generating faster and at lower cost. Specific claims from that page:

• Image editing at up to 4 megapixels across FLUX.2 variants.
• Support for up to 10 simultaneous reference images, which BFL describes as helping character, product and style consistency.
• Better handling of "complex typography, infographics, memes and UI mockups," which BFL says now work reliably in production.
• Improved adherence to complex, structured, multi-part prompts.

Those are the maker's own claims, so treat them as a description of intent, not a guarantee. Text rendering in particular still varies run to run on every model. If a sign has to be exact, generate a few versions and check every letter.

The same page lists other FLUX.2 variants: flex (developer-controlled quality and speed parameters), dev (a 32B open-weight model) and klein (listed there as coming soon). They aren't the same as Pro, so don't assume Pro's behaviour carries over.

Side by side

| | schnell | dev | FLUX 2 Pro |

|---|---|---|---|

| Generation | FLUX.1 | FLUX.1 | FLUX.2 |

| Open weights license | Apache 2.0 | Non-commercial | Pro is a hosted model |

| Built for | Speed | Quality over schnell | Production work |

| Reference images | Not a headline feature | Not a headline feature | Up to 10 (per BFL) |

| Best use | Drafts | Second pass | Final output |

A workflow that saves money and time

Don't start with the most capable model. Start with the cheapest one that answers your question.

1. Draft on schnell. Write a short prompt and run several variations. You're judging composition and mood only.
2. Lock the idea. Pick the version closest to what you want and note what the prompt got right.
3. Add detail to the prompt. Use a structure like the one in AI Image Prompt Structure: Subject, Setting, Light, Lens.
4. Run it on FLUX 2 Pro for the final, especially if it has text, a product, or several elements that must all appear.
5. Check the details before publishing: hands, letters, logos, edges.

An example prompt to try

Draft prompt on schnell:

Small coffee shop at night, neon sign in the window, rain on the street

Finishing prompt on FLUX 2 Pro:

Small corner coffee shop at night, rain-slick pavement, warm light glowing from the window, a neon sign reading "OPEN LATE" in clean block letters, 35mm lens, shallow depth of field, cinematic color

Notice what changed: the text is in quotes, the font style is described, and the lens and light are named. If you want more on that, see How to Get Readable Text Inside AI Images.

Common mistakes

• Judging a model on one image. Results vary run to run. Generate several before you decide a model is bad.
• Using schnell for a final with text. It's a draft tool. Move up a model when the words matter.
• Copying settings between models. schnell's example uses 4 steps and guidance 0.0; dev's example uses 50 steps and 3.5. They're not interchangeable.
• Piling on negative prompts. Not every model uses them. See Negative Prompts: When They Help and Who Ignores Them.
• Ignoring licensing. Open weights don't all carry the same terms. Read them if you're doing commercial work.

Trying them in one place

In Mediafy Studio you can pick Image, choose a model such as Seedream 5 Pro, GPT Image 2 or Nano Banana 2, and run the same prompt on each. Each model has its own credit cost, shown on the Generate button before you run it, so you can check the cost before a final. If a prompt is too short, the Enhance button rewrites it into a fuller one. If you'd like a more structured path through all of this, there's also the AI Creator Lab course.

What these models can't do

Be realistic. None of the three is a guarantee of correct text, correct anatomy or a faithful match to a brand. FLUX models don't know your product, and a reference image helps but doesn't make output identical. Always review the result before it goes out under your name.

Ready to compare them yourself? Sign in to Mediafy Studio and run one prompt across all three.

Sources

• Black Forest Labs, "FLUX.2" announcement: https://bfl.ai/blog/flux-2
• Black Forest Labs (Hugging Face), "FLUX.1-schnell" model card: https://huggingface.co/black-forest-labs/FLUX.1-schnell
• Black Forest Labs (Hugging Face), "FLUX.1-dev" model card: https://huggingface.co/black-forest-labs/FLUX.1-dev