FLUX.1: Stable Diffusion creators launch new image AI

Black Forest Labs introduced FLUX.1 on August 1, 2024, a family that turns written instructions into images. The company brought together researchers involved in VQGAN, latent diffusion, Stable Diffusion and Stable Diffusion XL. Its first release was not one model at three speeds, but three products with different access and rights: [pro] through an API, [dev] with weights for non-commercial use and [schnell] under Apache 2.0.

The distinction changes what an artist, researcher or company can do. One version may produce the preferred image while not allowing its weights to become a service. Another may support commercial integration and trade finish for speed. Before comparing galleries, adopters need a four-column sheet: quality, cost, control and licence.

Three suffixes that are part of the product

Black Forest Labs’ original announcement defined FLUX.1 [pro] as the highest-quality version, available through the company’s API and partners. A user sent a request and received an image without downloading weights. This route simplifies infrastructure and scaling but places data, availability and terms with the service provider.

FLUX.1 [dev] was a distillation of [pro] designed to retain quality and prompt adherence with greater efficiency. Its weights were downloadable for non-commercial applications. The official [dev] model card describes a 12-billion-parameter rectified-flow transformer trained through guidance distillation and distributed under a non-commercial licence.

FLUX.1 [schnell]—German for fast—was optimised to generate in a few steps. Its weights were released under Apache 2.0, a permissive licence that allows commercial use under its terms. In the original repository, the main weights file was 23.8GB: open access does not mean a small download or instant inference on every machine.

The official inference code runs the accessible-weight models. The repository’s Apache 2.0 licence does not change [dev]’s model licence: code and weights are separate artefacts. Copying the licence visible at the root without opening the model agreement can lead to a commercial deployment that was never authorised.

The model, output and API have different contracts

The initial [dev] licence allowed access, use, derivatives and distribution only for non-commercial purposes. The agreement associated with that version said commercial activity involving the model or a derivative required an additional licence from Black Forest Labs.

Yet the same text said the company claimed no ownership of outputs and allowed them to be used commercially, subject to the agreement’s prohibitions. It also barred using outputs to train, fine-tune or distil a competing model. “[dev] non-commercial” therefore described the weights and derivatives, not an absolute ban on selling a generated illustration. This is why the legal answer cannot fit into a page label.

[schnell] addressed commercial weight use differently: Apache 2.0 grants broad rights to use, modify and distribute, with duties such as preserving notices. [pro] added API terms, not the licence of downloadable weights. A proper audit opens the exact agreement for both the variant and access route; it never transfers permissions from one sibling to another.

Rights in inputs and outputs remain in addition to the model licence. A provider’s decision not to claim an image does not guarantee that the user owns every right in a brand, face or work named in a prompt. The model contract does not replace law or third-party permission.

What a flow model does

Black Forest Labs described the public models as a hybrid of multimodal and diffusion-transformer blocks, scaled to 12 billion parameters and built on flow matching. The system learns a path from a noise distribution to an image representation, conditioned on text. Generation follows that path in several steps.

A transformer helps connect words, positions and visual components; the flow process constructs an image from a disordered representation. This explains how [schnell] can be faster: distillation attempts to learn a route requiring fewer steps. Fewer steps reduce time, but the relationship with quality must be measured at the actual resolution and hardware.

Twelve billion parameters are not a visual grade. Architecture, data, post-training, decoder, sampler and step count all affect the result. Two variants of similar scale can differ greatly because one was distilled from another or optimises a different objective.

How to test the quality claim

Black Forest Labs said [pro] and [dev] surpassed Midjourney v6.0, DALL·E 3 HD and Stable Diffusion 3 Ultra in visual quality, prompt following, formats, typography and diversity. This was a vendor comparison. The announcement did not provide a complete protocol that turned its chart into a universal result for every style, language and application.

A useful evaluation begins with a fixed prompt bank. It should include simple and complex scenes, counts, spatial relations, exact text, hands, partial occlusion, several permitted styles and diverse people. A commercial catalogue adds background, brand colours, margins and negative space. The same set runs across every variant.

Several images should be generated per prompt with recorded seeds. Selecting only the best output measures curation ability, not one-generation reliability. Teams can run blind pairwise comparisons and record four rates: element compliance, severe defects, required retouching and actually usable output. Time, cost and memory belong beside them.

Typography needs a literal criterion: a sign that merely resembles text does not pass. Reviewers transcribe the result and compare it with the requested string. Spatial relations are scored condition by condition—object A left of B, three items, correct colour. Decomposing the prompt prevents an attractive image from hiding task failure.

Diversity is more than visual variety

The announcement highlighted output diversity. One test repeats a prompt that does not specify age, skin, body or setting and observes the distribution, without pretending a small sample represents training. It then adds explicit attributes and measures whether the model follows them without unsolicited stereotypes.

Reviewers should also look for duplication: compositions, faces or brands that recur too often. A creative tool can produce superficially different images while repeating one structure. Useful diversity combines variation across seeds with control when the user does specify a condition.

Training-data provenance was not detailed in the announcement. Accessible weights support inspection and adaptation, but they do not automatically provide a corpus inventory. A user who needs traceability for a campaign should record that absence as a limit, not treat architecture as an answer.

Choose for the workflow, not the podium

[pro] could suit users prioritising the best claimed result without operating GPUs. [dev] offered local experimentation, adaptation and process control within a non-commercial framework. [schnell] favoured prototypes, previews and commercial integrations where speed and a permissive licence outweighed finish. No choice was universal.

A pilot should reproduce the whole chain: input, generation, filtering, human review, editing, storage and publication. Public content also needs provenance records, a policy for real people and a signal when an image could be mistaken for documentation. A provider filter does not relieve the publisher of responsibility.

The transferable skill is to separate three objects before using any generator: model weights, generated images and the service running the model. Each may carry a different contract and risk. FLUX.1 expanded competition and local control, but the family only made sense with the [pro], [dev] and [schnell] suffixes intact. Removing them turned three technical and legal decisions into an imprecise brand.