Flux review

What it is

Flux is the image model family from Black Forest Labs, a team that includes several original Stable Diffusion researchers. It arrived in 2024 and quickly became the default open-weight choice, because it fixed the two things that most limited earlier open models: human anatomy and text.

Where it shines

Photorealism. Faces, hands and skin — the classic failure modes — are handled better than any other openly available model, and often better than the closed competition. For product photography, lifestyle imagery and anything that must pass as a real photograph, it is the strongest option that you can also run yourself.

Text rendering is the second win. Signage, packaging labels and short headlines come out legible, which for a long time was the exclusive advantage of closed models.

Because the lighter variants have open weights, Flux appears in a wide range of hosts and tools, and competition between them has driven per-image prices down to fractions of a cent. Fine-tuning support means studios can train it on their own products just as they did with Stable Diffusion.

Where it falls short

There is no polished first-party app. Black Forest Labs ships models, not products, so your experience depends entirely on whichever third-party interface you choose — and quality, defaults and prompt handling differ noticeably between them.

Licensing needs attention: the fast Schnell variant is permissively licensed, the higher-quality Dev variant is non-commercial without a separate licence, and the best variants are API-only. Getting this wrong is an easy mistake in commercial work.

Stylistically it is also less distinctive than Midjourney; it produces excellent photographs but is less likely to produce something surprising.

Pricing in plain English

Free if you self-host the open variants. From around $0.003 per image via API. Ten to thirty dollars a month through hosted interfaces that bundle editing tools.

Who should use it

Developers building image features into products, studios needing photoreal output at volume, and anyone who wants near-frontier quality with the option to run it on their own hardware.