Review draws on 5 primary sources (vendor announcements, named publications, benchmark results) and is updated continuously as the product changes. See the methodology page for the full research process.
TL;DR: FLUX is the model family from Black Forest Labs — the German lab founded by the researchers who created Stable Diffusion. Unlike Midjourney, FLUX isn’t an app you subscribe to: it’s the engine other products build on (Mistral’s Le Chat, Adobe Photoshop’s generative editing, Meta’s Vibes video app, NVIDIA’s Blackwell foundation models). The family runs from free Apache-2.0 open weights (FLUX.2 [klein], small enough for consumer GPUs) through a ~$0.03–0.045 per image API (FLUX.2 [pro]), with megapixel-based scaling. Update (July 23, 2026): BFL announced FLUX 3 — a single multimodal model for image, video up to 20 seconds with native audio, and action prediction. FLUX 3 Video and Action are in early access (video from $0.17/second in draft HD); FLUX 3 Image is rolling out over the coming weeks, and open-weight versions are promised later in 2026. There’s even a robotics variant, FLUX-mimic, that Audi is testing on production lines. Choose FLUX for ownership, editing precision, and typography; choose Nano Banana Pro for speed and zero cost; choose Midjourney for out-of-the-box aesthetics.
What FLUX is in 2026
FLUX comes from Black Forest Labs, the Freiburg-based lab that Robin Rombach and several co-authors of the original Stable Diffusion paper founded in 2024. That lineage matters: the team that made open-weights image generation mainstream once has spent two years doing it again, with better models and an actual business model this time.
The key thing to understand before comparing it to anything: FLUX is a model family, not a product. There is no polished FLUX consumer app with a prompt box and a subscription. You use FLUX in one of three ways:
- Through other products that embed it. Mistral’s Le Chat generates images with FLUX. Adobe Photoshop’s in-context generative editing runs FLUX.1 Kontext Pro. Meta’s Vibes video app was built in collaboration with BFL. If you’ve generated an image inside one of those, you’ve used FLUX without knowing it.
- Through the API (BFL’s own platform, or hosts like fal, Replicate, and Together) — pay per image, no subscription.
- On your own hardware. Several FLUX models ship as downloadable open weights. This is the option Midjourney and Google simply don’t offer.
That third path is FLUX’s identity. When Grok needed image generation at launch, it used FLUX. When NVIDIA wanted image foundation models optimized for Blackwell GPUs, it partnered with BFL. The open-weights ecosystem — ComfyUI workflows, LoRA fine-tunes, community checkpoints — that grew up around Stable Diffusion has largely migrated to FLUX.
The model family, decoded
BFL ships a lot of variants. Here’s the map, current as of August 2026:
| Model | What it is | License / access |
|---|---|---|
| FLUX.2 [pro] | The flagship image API model | API only, ~$0.03–0.045/image |
| FLUX.2 [max] | Higher-quality tier above pro (Dec 2025) | API only |
| FLUX.2 [flex] | Specialist for typography and text rendering | API only |
| FLUX.2 [dev] | Open weights, 32B-class, best self-hosted quality | Free for non-commercial; paid license for commercial |
| FLUX.2 [klein] | Distilled small models — 4B and 9B (Jan 2026) | Apache 2.0 — free, including commercially |
| FLUX.1 Kontext | In-context editing: modify images with text instructions | Max/Pro via API; Dev open (non-commercial) |
| FLUX.1 [schnell] | The 2024 fast model, still everywhere | Apache 2.0 |
| FLUX 3 | New multimodal generation (see below) | Early access |
Two practical notes. First, the license split is a genuine trap: [dev] weights are free to download but not free to use commercially — BFL sells open-weights licensing tiers (Builder, Platform, Professional, Enterprise) with monthly image allowances for that. If you’re generating images for anything that makes money and want to self-host free of charge, use [klein] (Apache 2.0) — that’s exactly what it exists for. Second, FLUX.2’s architecture pairs the image model with a Mistral 24B vision-language model for prompt understanding, which is why prompt adherence jumped so visibly over FLUX.1.
FLUX 3 — what actually launched on July 23
The reason FLUX is in the news: on July 23, 2026, BFL announced FLUX 3, its first frontier model trained jointly on image, video, audio, and action in one architecture. (Our launch-day coverage has the full breakdown.)
What that means in practice:
- FLUX 3 Video generates clips up to 20 seconds in a single generation, with native audio as an optional feature — text-to-video, image-to-video, video-to-video from a reference clip, keyframe-to-video for controlled transitions, and audio-video continuation. BFL claims it leads early evaluations against frontier video models, with particular strength in facial expressions and matching sounds to physical events. Treat vendor claims as claims until independent benchmarks land.
- FLUX 3 Image was announced but is not generally available yet — BFL says it rolls out “in the coming weeks,” with general availability after benchmarks are published.
- FLUX-mimic, built with Swiss startup mimic robotics, extends the same model to robot manipulation — fine-tunable for a new task with ~30 minutes of robot data instead of the 30+ hours prior approaches needed. Audi is testing it for soft-body manipulation in production and logistics. That’s not a consumer feature, but it tells you where BFL is aiming: one visual-intelligence stack from marketing images to factory floors.
- Open weights are promised again: faster, open-weight FLUX 3 versions are planned for later in 2026.
The honest read: as of early August 2026, FLUX 3 is an early-access product for video teams and an announcement for everyone else. If you came here deciding what to use today, the shipping products are the FLUX.2 family — which is mature — plus FLUX 3 Video if you get early access and need generated video with sound.
Pricing — what it actually costs
FLUX pricing is pay-as-you-go: no subscriptions, no seats.
- API images (FLUX.2 [pro]): roughly $0.03–0.045 per image, scaling with output megapixels and reference-image use. A thousand marketing images ≈ $30–45. For comparison, that undercuts most commercial image APIs.
- FLUX 3 Video (early access): from $0.17 per second in draft mode at HD — a 10-second draft clip ≈ $1.70. Final-quality renders cost more.
- Self-hosted, actually free: FLUX.2 [klein] (4B or 9B) and FLUX.1 [schnell] are Apache 2.0 — run them locally or on rented GPUs with no per-image cost and full commercial rights.
- Self-hosted, watch the license: FLUX.2 [dev] gives the best open quality but needs a paid BFL license for commercial use, sold in tiers with monthly image allowances.
- Hardware reality: [dev]-class models want a serious GPU (24GB+ VRAM, or quantized builds on less); [klein] 4B runs on ordinary consumer cards. If you don’t own a GPU, hosted APIs erase the difference.
What FLUX does well
Ownership and control. The only frontier-quality image family you can download, fine-tune, and run air-gapped. LoRA fine-tuning on your product catalog or art style is routine in the ComfyUI ecosystem — impossible with Midjourney or Nano Banana Pro.
Editing, not just generating. FLUX.1 Kontext made “change the jacket to red, keep everything else” work reliably — which is why Adobe put it inside Photoshop. Instruction-based editing remains a FLUX signature strength.
Typography. FLUX.2 [flex] exists specifically because text-in-image is the classic failure mode of image models. Posters, packaging mockups, UI concepts — this is the specialist tool.
Prompt adherence. The Mistral-24B language front-end in FLUX.2 understands long, structured prompts noticeably better than FLUX.1 did. Complex scene descriptions with spatial relationships mostly land.
An industry default. Mistral, Adobe, Meta, NVIDIA, and (early on) xAI all chose FLUX to build with. For anyone making a build-vs-buy decision, that adoption list is the strongest available signal that the models and licensing hold up in production.
Where FLUX falls short
It’s not an app. No native consumer experience with community feeds, style references, and a Discord full of prompt inspiration. Midjourney’s product polish, and the sheer convenience of Nano Banana Pro sitting inside Gemini, serve casual users far better. FLUX assumes you’ll bring the interface.
The license maze. Schnell and klein Apache, dev non-commercial, Kontext split by tier, pro API-only — it takes a table (see above) to know what you’re allowed to ship. Midjourney is one subscription; Nano Banana Pro is free. Simplicity counts for something.
Aesthetic ceiling out of the box. Untuned FLUX output is excellent but reads neutral; Midjourney still wins default artistic “wow” per prompt. FLUX catches up when you fine-tune — which most casual users never will.
FLUX 3 is mostly promise today. Video and Action in early access, Image “coming weeks,” open weights “later in 2026.” BFL has shipped on schedule before, but today you’re mostly buying the FLUX.2 family plus a roadmap.
Churn risk in integrations. xAI’s Grok started on FLUX and later moved to its own Aurora model. Being the engine inside other products means the products can swap engines.
FLUX vs the alternatives
For out-of-the-box beauty: Midjourney > FLUX. Midjourney’s default aesthetic and community remain unmatched for art-direction-free results.
For free and fast: Nano Banana Pro > FLUX. Google’s model inside Gemini generates photorealistic images in 1–3 seconds at $0. FLUX can’t beat free-and-instant for casual use.
For self-hosting, fine-tuning, and commercial control: FLUX > everything. This is the category it owns; Stable Diffusion is its spiritual predecessor and the ecosystem has largely moved over.
For instruction-based editing: FLUX (Kontext) ≈ Nano Banana Pro > Midjourney. If your workflow is “edit this image,” those two lead.
For text in images: FLUX.2 [flex] > the field.
For generated video with audio: early days — FLUX 3 Video is a genuine contender by claim, but Runway is the shipping professional tool; wait for independent benchmarks.
Full category ranking: best AI image-generation tools.
Who should use FLUX
- Developers and product teams embedding image generation via API at $0.03/image
- Studios and agencies that need fine-tuned, on-brand, self-hosted generation
- Anyone whose images must carry legible text — [flex] is the specialist
- Privacy-bound or air-gapped environments — klein/dev run fully offline
- Video-generation early adopters with FLUX 3 early access
Who shouldn’t
- Casual creators — Nano Banana Pro (free, instant) or Midjourney (best defaults) will make you happier
- One-subscription simplicity seekers — the license/variant maze isn’t worth it for occasional images
- Teams that need shipping video today — use Runway until FLUX 3 exits early access
The verdict
FLUX in 2026 is the infrastructure choice in image generation: the family you pick when you need ownership — of weights, of fine-tunes, of costs, of deployment. The people who built the open-weights image era at Stability rebuilt it properly at Black Forest Labs, and the adoption list (Adobe, Mistral, Meta, NVIDIA) shows it. With FLUX 3, BFL is betting the same playbook extends from images to video, audio, and even robot arms — and unusually for this market, it has earned some benefit of the doubt.
The pragmatic split: if you build things, FLUX belongs on your shortlist; if you just want great images, it’s the engine — go use one of the cockpits instead.
Related:
Sources: Black Forest Labs FLUX 3 announcement (GlobeNewswire, Jul 23, 2026) · bfl.ai pricing · FLUX — Wikipedia · FLUX API pricing breakdown (Puter, Jun 2026) · Price Per Token — FLUX pricing
FLUX — frequently asked questions
What does FLUX do?
FLUX comes from Black Forest Labs, the Freiburg-based lab that Robin Rombach and several co-authors of the original Stable Diffusion paper founded in 2024. That lineage matters: the team that made open-weights image generation mainstream once has spent two years doing it again, with better models and an actual business model this time. The key thing to understand before comparing it to anything: FLUX is a model family, not a product. There is no polished FLUX consumer app wi…
How much does FLUX cost?
FLUX pricing is pay-as-you-go: no subscriptions, no seats. - API images (FLUX.2 [pro]): roughly $0.03–0.045 per image, scaling with output megapixels and reference-image use. A thousand marketing images ≈ $30–45. For comparison, that undercuts most commercial image APIs. FLUX 3 Video (early access): from $0.17 per second in draft mode at HD — a 10-second draft clip ≈ $1.70. Final-quality renders cost more. Self-hosted, actually free: FLUX.2 [klein] (4B or 9B) and FLUX.1 [sch…
What are the downsides of FLUX?
It's not an app. No native consumer experience with community feeds, style references, and a Discord full of prompt inspiration. Midjourney's product polish, and the sheer convenience of Nano Banana Pro sitting inside Gemini, serve casual users far better. FLUX assumes you'll bring the interface. The license maze. Schnell and klein Apache, dev non-commercial, Kontext split by tier, pro API-only — it takes a table (see above) to know what you're allowed to ship. Midjourney is…
What are the best alternatives to FLUX?
For out-of-the-box beauty: Midjourney > FLUX. Midjourney's default aesthetic and community remain unmatched for art-direction-free results. For free and fast: Nano Banana Pro > FLUX. Google's model inside Gemini generates photorealistic images in 1–3 seconds at $0. FLUX can't beat free-and-instant for casual use.
Who should use FLUX?
Developers and product teams embedding image generation via API at $0.03/image Studios and agencies that need fine-tuned, on-brand, self-hosted generation Anyone whose images must carry legible text — [flex] is the specialist Privacy-bound or air-gapped environments — klein/dev run fully offline Video-generation early adopters with FLUX 3 early access
Is FLUX worth it in 2026?
FLUX in 2026 is the infrastructure choice in image generation: the family you pick when you need ownership — of weights, of fine-tunes, of costs, of deployment. The people who built the open-weights image era at Stability rebuilt it properly at Black Forest Labs, and the adoption list (Adobe, Mistral, Meta, NVIDIA) shows it. With FLUX 3, BFL is betting the same playbook extends from images to video, audio, and even robot arms — and unusually for this market, it has earned so…
Thinking about trying FLUX?
The button below goes to FLUX's official site. Signing up through it may earn this site a small commission at no cost to the reader. That helps keep Pick Right running and is never the reason a tool gets recommended.
Learn More →