FLUX is a family of text-to-image generation and image editing models developed by Black Forest Labs, a German company founded in 2024 by researchers who previously worked on latent diffusion and Stable Diffusion. Among AI models & assistants, FLUX stands out as a rectified flow transformer that is released in several tiers, from closed API models to downloadable open weights.
The first generation, FLUX.1, appeared in August 2024 and quickly became a reference point for prompt adherence, legible text in images and realistic human anatomy. Later releases added faster API models, control and inpainting tools, the in-context editing family FLUX.1 Kontext and a second generation, FLUX.2, which extends multi-reference editing.
How FLUX generates images
FLUX generates images in a compressed latent space rather than directly in pixels. A variational autoencoder encodes images into a latent grid with more channels than earlier Stable Diffusion models, which preserves fine texture and small text after decoding. The generative model then learns to transform random noise into such latents, conditioned on the text prompt.
The training objective is rectified flow, a form of flow matching. Instead of learning to remove noise step by step along a curved diffusion path, the model learns a velocity field along nearly straight paths between noise and data. Straighter paths need fewer integration steps at sampling time, which is why distilled FLUX variants can produce usable images in one to four steps.
The backbone of FLUX.1 is a transformer of about 12 billion parameters that combines double-stream blocks, where text and image tokens keep separate weights but attend to each other, with single-stream blocks that process the joined sequence. Text conditioning comes from a T5 text encoder for detailed semantics and a CLIP encoder for a pooled global signal, and rotary positional embeddings let the model handle different resolutions and aspect ratios.
The table below lists the main architectural components of FLUX.1 and their role.
| Component | Role | Practical effect |
|---|---|---|
| Latent autoencoder | Compresses images into latent grids | Sharper detail and readable text |
| Rectified flow objective | Learns straight noise-to-image paths | Fewer sampling steps |
| Double-stream blocks | Joint attention with separate weights | Strong prompt-image alignment |
| Single-stream blocks | Process merged text and image tokens | Efficient deep refinement |
| T5 and CLIP encoders | Turn prompts into conditioning signals | Long, detailed prompts are followed |
| Rotary embeddings | Encode 2D token positions | Flexible resolution and aspect ratio |
FLUX model variants and licenses
FLUX is released in tiers that differ in speed, quality and license, and the license often decides which variant a team can use. The open variants are produced by distillation from the strongest internal model: guidance distillation removes the need for classifier-free guidance passes, and timestep distillation compresses sampling to a few steps.
| Variant | Access | License | Typical steps |
|---|---|---|---|
| FLUX.1 [pro] | Hosted API only | Commercial API terms | About 25 to 50 |
| FLUX.1 [dev] | Open weights | Non-commercial weights license | About 20 to 50 |
| FLUX.1 [schnell] | Open weights | Apache 2.0 | 1 to 4 |
| FLUX1.1 [pro] | Hosted API only | Commercial API terms | Optimized for speed |
Outputs generated with [dev] can generally be used commercially, but hosting or fine-tuning the [dev] weights for a commercial service requires a separate license. The [schnell] model is the only first-generation variant under a permissive open-source license, which made it popular for local tools and products.
Later releases broadened the family beyond plain generation. The table below summarizes them; exact parameter counts and licenses vary by release and should be confirmed before deployment.
| Release | Purpose | Key capability |
|---|---|---|
| FLUX.1 Fill | Inpainting and outpainting | Edits masked regions consistently |
| FLUX.1 Depth and Canny | Structural control | Follows depth maps or edge maps |
| FLUX.1 Redux | Image variation | Mixes a reference image with a prompt |
| FLUX.1 Kontext | In-context editing | Edits images from text instructions |
| FLUX.2 | Second-generation generation and editing | Multiple reference images per request |
How FLUX compares with other image models
FLUX competes with closed hosted models such as Imagen, Adobe Firefly, Ideogram and the image features of large assistants, and with open models such as Stable Diffusion 3.5. Its main distinction is the combination of frontier-level quality with open weights in at least one tier, which allows local inference, fine-tuning and custom pipelines.
| Model family | Weights | Typical strength |
|---|---|---|
| FLUX | Open and closed tiers | Prompt adherence and fine-tuning ecosystem |
| Stable Diffusion 3.5 | Open weights | Broad tooling and lower hardware needs |
| Imagen | Closed, hosted | Photorealism within a large cloud platform |
| Adobe Firefly | Closed, hosted | Licensed training data for commercial design |
| Ideogram | Closed, hosted | Typography and graphic layouts |
Rankings change with each release, so comparisons should rely on current blind preference tests and in-house prompts rather than on reputation. Quality differences between leading models are often smaller than differences caused by prompt wording and sampling settings.
Running and adapting FLUX in practice
Open FLUX weights run in common diffusion toolkits and node-based interfaces. At full 16-bit precision the 12-billion-parameter transformer plus text encoders needs a GPU with roughly 24 GB of memory, while 8-bit or 4-bit quantization and CPU offloading make generation possible on consumer cards at the cost of speed and some fidelity.
- LoRA fine-tuning adapts FLUX to a person, product or style with a few dozen curated images and modest compute.
- Guidance scale for [dev] is a distilled parameter; values around 3 to 4 are common, and high values oversaturate images.
- Natural-language prompts work better than tag lists, because the T5 encoder reads full sentences.
- Fixed seeds and recorded settings are essential for reproducible results across a project.
Limitations and risks of FLUX
FLUX still shows the typical weaknesses of image generators: inconsistent counting of objects, errors in complex hands or crowds, and occasional distorted text in long phrases. Distilled variants trade some diversity for speed, and many users observe a recognizable default aesthetic, such as smooth skin and shallow depth of field, that needs prompting or fine-tuning to avoid.
Open weights also create governance responsibilities. Hosted APIs apply content filters and provenance metadata, but local deployments must add their own safety filtering, consent checks for likeness training and output labeling. Questions about training data provenance remain open, so organizations with strict compliance requirements should review the current license and data statements before adoption.
Put simply
FLUX is a set of AI tools that turn a written description into a picture. It can also change an existing picture based on instructions. It comes in several versions. Some are paid online services. Others can be downloaded and run on a strong home computer.
Think of it as a skilled illustrator who has studied millions of images. It can draw almost anything you describe, even readable signs and labels. The main idea is choice: a fast free version for experiments, a more careful version for quality, and paid online versions for businesses.
In practice
A small furniture retailer needs lifestyle photos for 300 products but cannot afford a studio shoot for every item. The design team photographs each product once on a plain background, trains a LoRA adapter on about 30 images of the brand’s interiors, and uses an open FLUX model on a single 24 GB workstation GPU to generate room scenes around each product.
An editing model then places the real product photo into the generated scene, so the item itself is never redrawn. Every image passes a human check for scale, shadows and accurate colors before publication. The team produces a full catalog of scenes in about two weeks and keeps product appearance faithful to the real items, which avoids misleading customers while cutting photography costs.