FLUX Review 2026
Open-source photorealism from the creators of Stable Diffusion — a 12B parameter model matching Midjourney-quality output, free self-hosted or available by API, with FLUX.2 Kontext for advanced multi-reference image editing.
FLUX Review Verdict — March 2026
What FLUX actually does
Open-weight models you can run locally or via API. Three core tiers support everything from fast prototyping to production-grade output, while FLUX.2 Kontext adds advanced image editing control.
Available FLUX models
FLUX Pricing — March 2026
Free self-hosted access for FLUX.1 [schnell] and FLUX.1 [dev], API pricing for hosted production usage, and commercial licensing options for enterprise workflows.
| Model | Price | License | Access | Commercial use | Best for |
|---|---|---|---|---|---|
| FLUX.1 [schnell]Free | Free Self-hosted |
Apache 2.0 | GitHub, Hugging Face | ✓ Full | Fast prototyping |
| FLUX.1 [dev] | Free Self-hosted |
Non-commercial | GitHub, Hugging Face | License req. | Research & personal |
| FLUX.1 [pro] | ~$0.05/img API pricing |
Proprietary | BFL API, Replicate | ✓ Full | Production work |
| FLUX.2 [klein] | $0.014/img Sub-second |
Apache 2.0 | API + Self-hosted | ✓ Full | High volume |
What works and what doesn’t
FLUX offers elite open-source value and control, but the trade-off is clear: you need more hardware, more technical skill, and a more hands-on workflow than web-based AI image tools.
✓ Strengths
FLUX stands out because it combines serious image quality with open access, flexible deployment, and zero recurring costs for users who can run it locally.
With Apache 2.0 licensing, you can self-host schnell with zero recurring costs and unlimited generation on your own hardware.
Ars Technica testing found FLUX output quality comparable to leading closed models, especially for realistic image generation.
FLUX performs significantly better than Stable Diffusion XL when generating consistent and realistic human hands.
Robin Rombach, Andreas Blattmann, and Patrick Esser bring proven expertise from one of the most influential open-image-model projects in the market.
You can self-host locally, build workflows in ComfyUI, or use hosted APIs such as Replicate and fal.ai depending on your needs.
Black Forest Labs raised $31M, including funding from Andreessen Horowitz, which supports continued innovation and product expansion.
✗ Weaknesses
FLUX is powerful, but it is not beginner-first. The biggest limitations come from setup complexity, hardware demands, and fewer polished consumer-facing workflows.
Running local inference typically requires GPU hardware, Python tooling, and CLI or workflow setup knowledge, so it is not ideal for beginners.
With 12B parameters, FLUX works best with 16GB+ VRAM. Many consumer GPUs can run it, but performance and convenience are lower.
Unlike Midjourney or DALL-E, FLUX does not center around a polished web product. Many users will need ComfyUI or another workflow layer.
Black Forest Labs has not fully disclosed its exact training data sources, which raises the same copyright and provenance questions found across the broader market.
The dev model is free for personal and research use, but commercial deployment requires separate paid licensing from Black Forest Labs.
Very realistic outputs can enable deceptive or unethical use cases, so FLUX faces some of the same criticism directed at other advanced photorealistic models.
FLUX Review FAQ
Yes. FLUX.1 [schnell] is completely free under the Apache 2.0 license, including commercial use. FLUX.1 [dev] is free for non-commercial use. For commercial use of [dev] or access to [pro], you need licensing or paid API access from Black Forest Labs. Self-hosting removes ongoing API costs entirely.
According to Ars Technica testing, FLUX photorealism closely matches Midjourney 6, while prompt fidelity is comparable to DALL-E 3. FLUX also performs especially well on hand rendering. The main trade-off is usability: Midjourney has a simpler interface, while FLUX offers open weights, self-hosting, and more technical control.
FLUX has 12 billion parameters, so local usage benefits from strong GPU memory. A practical recommendation is 16GB+ VRAM. FLUX.1 [schnell] is faster and works better on consumer hardware than [dev], especially because it needs only 1–4 steps rather than a much heavier inference cycle.
FLUX was created by Black Forest Labs, founded in 2024 by Robin Rombach, Andreas Blattmann, and Patrick Esser — former Stability AI researchers who previously helped build Stable Diffusion while working at LMU Munich. The company has raised $31M, including backing from Andreessen Horowitz.
Yes, but it depends on the model. FLUX.1 [schnell] allows full commercial use under Apache 2.0. FLUX.1 [dev] is free for personal and research use, but commercial use requires a license from Black Forest Labs. FLUX.1 [pro] supports commercial use via paid API access.
FLUX.2 Kontext is the latest model tier that combines text-to-image generation with multi-reference image editing. It can use up to 10 reference images at once, letting you control colors, poses, composition, and object-level edits much more precisely through natural language instructions.
Open-source photorealism — run it free
12B parameters. Midjourney-quality output. Apache 2.0 licensing. Zero API costs when self-hosted.
Get FLUX on GitHub