The three variants, and which one you are actually downloading
Choosing “FLUX” without choosing a variant is the mistake this page exists to prevent.
| schnell | Apache 2.0. Distilled for speed, fewer steps, lower fidelity. Commercially usable with no further permission, no revenue threshold and no conditions. |
| dev | Non-commercial licence. The higher-quality open-weights variant. Free to run and experiment with; a separate commercial licence must be obtained for commercial output. |
| pro | API only. The strongest variant. No weights are published, so it cannot be run locally at all. |
dev is the one most people download, because it is the best model whose weights exist, and it is the one they may not use commercially. That single fact is the most consequential thing on this page.
The practical decision tree is short. Commercial work with no budget for a licence: schnell. Commercial work with budget: dev under a commercial licence, or pro through the API. Personal or research work: dev, and enjoy it.
Why prompt adherence improved
The architecture differs from the Stable Diffusion lineage — a rectified flow transformer rather than the older U-Net diffusion approach — and the practical effect is that instructions survive better.
Multiple subjects stay distinct. Spatial relationships hold: “the red cup to the left of the blue book” produces that arrangement more often than not. Attributes attach to the right objects rather than bleeding between them.
Those are exactly the failures that make image models frustrating for specification-like briefs, and improving them matters more for commercial work than another increment of photorealism.
Text inside images
Legible, correctly spelled text became realistic here rather than a lottery. Short strings — a sign, a label, a headline — come out right often enough to be usable.
It is not at Ideogram‘s level, and longer passages still degrade. But for an open-weights model it removed a limitation that had made local generation unusable for a whole class of design work.
What it costs in hardware
- For local use, a substantial GPU. The open variants are large — 12 GB VRAM is a realistic floor and 24 GB is comfortable. Quantised community builds lower that at some cost in quality and are widely used.
- An interface: ComfyUI has the strongest support, and other local tools have followed.
- For pro, an API key and per-image billing — no hardware at all.
- A licence decision made before you ship anything. This is a prerequisite, not an afterthought.
The VRAM requirement is a real step up from SD 1.5 and is why plenty of people running Stable Diffusion happily cannot run FLUX dev at full precision.
A younger ecosystem
Fine-tunes, LoRAs and ControlNet equivalents exist and are growing, but the depth is not comparable to Stable Diffusion‘s years of accumulation.
If your work depends on a specific community style model, check whether it exists for FLUX before switching. For general work the base model’s quality often removes the need; for niche styles it does not.
Training a LoRA on FLUX is possible and is more demanding than on SD 1.5, both in VRAM and in time.
What it does not cover
FLUX generates still images. It is not an editor, not an interface, and not an upscaler — pair it with ComfyUI for pipelines and Magnific or similar for enlargement.
Training-data provenance is not disclosed in detail, so it does not answer the question Firefly answers. Apache 2.0 on schnell governs the weights, not the material they learned from — a distinction worth being precise about if the question ever reaches a legal team.
Who should use it
People needing open weights and better instruction-following than Stable Diffusion offers — product visualisation, design work with specific requirements, anything where “close enough” is not.
Developers building generation into a product: schnell is one of the few genuinely unencumbered commercial options at this quality, and pro covers the top end without hardware.
Anyone needing legible text in generated images without using a hosted service.
It is a weak fit for someone with an 8 GB card, for anyone dependent on the large SD LoRA library, and for work where the licensing distinction cannot be managed carefully.
What the architecture delivers
- Prompt adherence at open-weights level — complex instructions survive.
- Usable text rendering, historically the clearest failure of open models.
- A genuinely permissive option in schnell: Apache 2.0, no thresholds, no conditions.
- A path from local to hosted without changing model family as workloads grow.
- Active development from a team with direct lineage in this architecture.
The trade-offs
- The three-licence structure catches people out, and dev is the trap.
- Heavy VRAM requirements relative to older Stable Diffusion models.
- A smaller ecosystem — fewer fine-tunes, LoRAs and workflows.
- pro is closed, so “open” only partly describes the family.
- Provenance is undisclosed, so it does not solve the indemnity question.
Your weights stay yours
For schnell and dev, the weights are files on your disk and keep working regardless of what the company does — subject to the licence you accepted when you downloaded them.
For pro there is nothing to keep. It is an API, and if access or pricing changes, your pipeline changes with it. That asymmetry is worth weighing when deciding which variant to build on: schnell is permanent in a way pro is not.
Models to weigh against it
- Stable Diffusion — far larger ecosystem, lower hardware demands, weaker adherence.
- Midjourney — better images out of the box, no weights, no control.
- Ideogram — hosted, and still the reference for text inside images.
- Adobe Firefly — the choice when indemnification outranks openness.
Compiled from Black Forest Labs’ documentation and public sources. Licence terms differ by variant — verify against the specific model you intend to use. We have not hands-on tested this tool. Last reviewed 16 August 2026.
