FLUX

FLUX is the model that made open-weights image generation credible against the closed leaders, and it did so by being genuinely good rather than merely available.

What open weights buy you

Three things that no hosted image API can offer, and teams usually want at least one of them.

Self-hosting: the images and prompts never leave your infrastructure. For anything involving unreleased products, client material under NDA, or personal likenesses, this is often the only acceptable arrangement.

Fine-tuning: train on your own material — a product catalogue, a brand style, a character — and generate consistently within it. LoRA training on FLUX is well-supported and comparatively cheap, and a well-trained LoRA outperforms prompt engineering against a general model for any repeated subject.

Cost at volume: once you are running your own inference, the marginal cost of an image is electricity. For applications generating thousands of images, this changes the economics entirely.

The variant question

FLUX ships in several variants with different licences and different capabilities, and getting this wrong is the most common mistake teams make.

The distilled fast variant is permissively licensed and quick, and is the right default for high-volume or interactive use. The higher-quality open variant carries a non-commercial licence, which rules it out for a substantial set of uses regardless of how good the output is. The top-tier model is available through hosted APIs rather than as downloadable weights.

Read the licence for the specific variant before you build on it. “FLUX is open source” is true in the same loose way that a lot of things are, and the details determine whether your product is legal.

Strengths and gaps

Prompt adherence is the standout. FLUX follows compositional instructions — this element here, that one there, in this relationship — more faithfully than the aesthetically-tuned competitors, which makes it better for work where the image has a specification rather than a vibe.

Text rendering is markedly better than most of the field, which removes one of the standing reasons to composite type separately.

Human anatomy is strong, and the classic generated-image tells are much less frequent.

The gaps: the default aesthetic is more neutral than Midjourney’s, which is a strength for specification-driven work and a weakness when you want the model to have taste on your behalf. The tooling ecosystem, while healthy, assumes competence with ComfyUI or a similar node-based pipeline. And self-hosting means owning GPU capacity, which is a real infrastructure commitment rather than a checkbox.

Pricing

Open weights are free to download for the variants that permit it, with cost shifting to hardware. Hosted endpoints are billed per image and are competitive with the closed alternatives.

The sensible pattern for most teams is to prototype on a hosted endpoint, measure whether volume justifies infrastructure, and self-host only if the arithmetic clearly works. Standing up GPU inference to save on a workload that generates two hundred images a month is a hobby, not a cost saving.

Who should choose it

Developers building image generation into a product, teams with data-residency constraints, anyone who needs a fine-tuned model for a consistent subject, and applications generating at volume.

Designers who want the best default aesthetic with no infrastructure should use Midjourney. Enterprise teams who need indemnification should use Adobe Firefly.

See best AI image generators for the comparison across all three.