Flux Review (2026): Black Forest Labs' Image Models, Pricing, and How They Compare

Flux Review (2026): Black Forest Labs' Image Models, Pricing, and How They Compare

Flux (styled FLUX) is one of the most capable AI image systems available in 2026, but it is easy to misunderstand what it is. Flux is not a single app you sign up for. It is a family of image models built by Black Forest Labs (BFL), a German lab founded by several of the researchers who originally created Stable Diffusion. You reach Flux through many different front-ends: BFL's own playground and API, third-party clouds like fal.ai and Replicate, no-code creative apps, or your own GPU. This hand-reviewed breakdown covers the current model lineup, where to run it, what it costs, and how it stacks up against Midjourney, DALL-E, and Stable Diffusion.

What is Flux?

Flux is a family of text-to-image and image-editing models based on a flow-matching transformer architecture. The single most important thing to grasp before choosing it: Flux is a model, not a product. There is no one "Flux app" the way there is a Midjourney bot or a ChatGPT window. Instead, the same underlying Flux models show up inside hosted playgrounds, dozens of third-party creative tools, cloud generation APIs, and local installs. That distribution model is Flux's defining trait and shapes almost everything about pricing and access below.

As of August 2026 the lineup spans three generations:

• FLUX.1 (2024). The original release, with [pro], [dev], and [schnell] variants, later joined by the faster, higher-quality FLUX1.1 [pro] and a FLUX1.1 [pro] Ultra mode for maximum resolution and photorealism. Still widely deployed and cheap.
• FLUX.2 (November 2025). The current mainstream generation and the one most people mean by "Flux" today. It ships as [pro], [flex], [dev], and [klein].
• FLUX 3 (announced July 2026). A multimodal frontier model that jointly learns from images, video, and audio in one architecture, extendable to robotics/action. As of this review, FLUX 3 Video and Action are in limited early access and FLUX 3 Image is still rolling out, so it is not yet a general-availability image tool.

The FLUX.2 lineup

FLUX.2 is where a 2026 buyer should focus. The four variants trade off openness, speed, and quality:

• FLUX.2 [pro] - the hosted flagship, tuned for the best balance of speed and quality for production text-to-image and image editing. API only.
• FLUX.2 [flex] - a hosted variant that exposes more control over steps and guidance, letting you trade compute for quality.
• FLUX.2 [dev] - a 32-billion-parameter open-weight model released under the FLUX Non-Commercial License. Weights are on Hugging Face and inference code on GitHub, but it is local-only (no hosted BFL endpoint) and needs a separate agreement for commercial use.
• FLUX.2 [klein] - distilled small models built for speed. The 4B version was released under a permissive Apache 2.0 license in January 2026, meaning it is free to use commercially on your own hardware; a larger 9B version ships under the non-commercial license. Klein targets sub-second generation on consumer GPUs.

Across the FLUX.2 family, the standout upgrades over FLUX.1 are multi-reference conditioning (feed up to 10 reference images to keep a character, layout, or style consistent), output and editing at up to 4 megapixels, noticeably sharper in-image text, and hex-color steering for brand-accurate palettes. Generation and editing live in the same checkpoint, so you no longer maintain separate models for the two jobs.

At a glance and pricing

Because Flux is delivered through many front-ends, there is no single price. You either pay per image through an API, pay nothing but hardware costs by self-hosting an open-weight variant, or fold it into a third-party app subscription. Representative 2026 figures:

On BFL's own API, FLUX.2 cost scales with output resolution: [pro] bills roughly $0.03 for the first megapixel of output and about $0.015 per megapixel after that, with reference-image inputs charged around $0.015 per megapixel. Third-party clouds can be cheaper for small images - Replicate and fal advertise generations from about $0.003, though Replicate uses per-second GPU billing so complex prompts cost more and budgets are less predictable. Older FLUX.1 [schnell] can run under a cent per image. These numbers come from BFL documentation and third-party pricing trackers checked in mid-2026; confirm against BFL's live calculator before budgeting, since per-megapixel rates change with resolution and inputs.

Key features and capabilities

Open weights as a first-class option. Unlike Midjourney or DALL-E, Flux gives you genuinely downloadable models. The [dev] and [klein] variants can run in ComfyUI or your own pipeline, and the Klein 4B model is commercially licensed. This is the single biggest reason builders choose Flux.

Photorealism. Flux is consistently rated among the best models for realistic human subjects, product photography, and architectural visualization, with 2026 comparisons placing FLUX.2 at or above Midjourney v7 and DALL-E for lifelike output.

In-image text. Reliable text rendering - fonts, kerning, multi-line layouts - is a genuine strength and a common reason to pick Flux for posters, mockups, and UI concepts.

Reference-driven consistency and editing. Multi-reference conditioning (up to 10 images) plus unified generation/editing make Flux strong for keeping a character or product consistent across a set, and for instruction-based edits.

Broad ecosystem. The same models are wired into fal.ai, Replicate, Together, OpenRouter, and consumer apps like Freepik, Krea, and Leonardo, so you are rarely locked to one vendor.

Output quality and control

Out of the box, Flux leans photoreal and technically precise rather than stylized. It tends to follow prompts literally - closer to DALL-E's obedience than Midjourney's opinionated aesthetic - which is what power users want when a brief has hard constraints. The [flex] variant and self-hosted [dev] expose step counts, guidance, and sampler settings for fine control, and the reference system gives repeatable results across a series. Anatomy and hands, historically a weak point for image models, are handled well.

The trade-off is that Flux does not impose the ready-made cinematic polish Midjourney is famous for. Casual users who want a gorgeous image from a three-word prompt may find Midjourney more satisfying with less effort; Flux rewards prompt craft and post-generation control.

Pros and cons

Pros

• Genuinely open weights ([dev], and commercially licensed [klein] 4B) for self-hosting and customization.
• Top-tier photorealism for people, products, and architecture.
• Reliable in-image text rendering.
• Multi-reference conditioning and unified generation + editing.
• Available across many APIs and apps, so no single-vendor lock-in.
• Low per-image cost, especially on third-party clouds and self-hosting.

Cons

• No single official app; beginners must pick a front-end, which adds friction.
• The best open model ([dev], 32B) is non-commercial and heavy to run locally.
• Licensing is fragmented across variants (Apache 2.0 vs. non-commercial vs. API terms).
• Less out-of-the-box artistic "wow" than Midjourney for casual prompts.
• FLUX 3 image generation is not yet generally available, so the frontier tier is still gated.

How it compares

• vs. Midjourney. Midjourney remains the aesthetic gold standard for cinematic, gallery-ready art via its $10-$60/month subscription and polished web app, but it offers no API, no open weights, and sometimes overrides your prompt for its own sense of beauty. Flux wins on control, prompt adherence, self-hosting, and per-image economics. See the full Flux vs Midjourney breakdown.
• vs. DALL-E / GPT Image. OpenAI's image models are strong all-rounders with excellent prompt adherence and easy ChatGPT access, but they are closed and API-metered. Flux matches or beats them on photorealism and text while adding an open-weight path.
• vs. Stable Diffusion. Both are open-weight and self-hostable, and they even share creators. Stable Diffusion has the larger fine-tune and LoRA ecosystem and runs on modest hardware, so it remains the pick for deep customization on a budget. Flux generally delivers higher base-model quality with less tuning.

In short: Midjourney for effortless art, DALL-E for conversational convenience, Stable Diffusion for the deepest customization, and Flux for the best mix of realism, control, and openness.

Who is Flux best for?

Flux is aimed squarely at builders and power users. Developers embedding image generation in a product, teams that need brand-consistent output at scale, and technical creators who want to self-host or fine-tune will get the most from it. The open [klein] 4B model is a rare commercially usable, self-hostable option, and the pro API is priced for volume. Casual users who just want one beautiful image with minimal setup are usually better served by Midjourney or DALL-E.

Verdict

Flux is the strongest open-leaning image system of 2026 and, for many photoreal and text-heavy jobs, the best available regardless of licensing. Its real value is optionality: pay-per-image APIs when you want managed infrastructure, open weights when you want control, and the same model quality either way. The catch is that Flux asks you to assemble your own workflow - choosing a front-end, navigating variant licenses, and crafting prompts - rather than handing you a single polished app. For builders and power users that is a feature; for casual creators it is friction. Judged as what it is, a model family rather than a consumer app, Flux earns a clear recommendation for anyone who needs realistic, controllable, and deployable image generation.

FAQ

Is Flux an app I can just sign up for? Not really. Flux is a family of models from Black Forest Labs. You use it through BFL's playground and API, third-party clouds like fal.ai and Replicate, consumer apps, or by self-hosting the open-weight versions.

What is the current Flux model? FLUX.2 (released November 2025) is the mainstream generation, in [pro], [flex], [dev], and [klein] variants. FLUX 3, announced July 2026, is a multimodal frontier model still in limited early access.

Is Flux open source? Partly. FLUX.2 [dev] (32B) is open-weight under a non-commercial license, and FLUX.2 [klein] 4B is Apache 2.0 (free for commercial use). The [pro] and [flex] tiers are closed and API-only.

How much does Flux cost? On BFL's API, FLUX.2 [pro] starts around $0.03 per megapixel and scales with resolution; third-party clouds can be cheaper per image, and self-hosting an open-weight variant costs only your hardware and electricity. Verify current rates against BFL's live pricing calculator.

Is Flux better than Midjourney? For photorealism, in-image text, prompt control, and self-hosting, generally yes. Midjourney still leads on effortless cinematic aesthetics and ease of use for casual creators.

Can I run Flux on my own computer? Yes, using the open-weight [dev] or [klein] models via tools like ComfyUI, subject to each variant's license and your GPU's memory.