Can You Legally Sell AI Art From FLUX? Licences Compared

FLUX.1 [dev] and FLUX.2 [dev] weights are under the FLUX Non-Commercial Licence — you cannot legally run them as part of a revenue-generating pipeline without buying a licence from Black Forest Labs. FLUX.1 [schnell], Qwen-Image, Qwen-Image-Edit-2509 and Z-Image-Turbo are all Apache-2.0 and carry no such restriction. That single split is what catches most people, because every tutorial on the internet uses [dev].

The awkward part, and the reason forum threads on this never resolve: the FLUX [dev] licence also says "You may use Output for any purpose (including for commercial purposes)." Both clauses are really in there. This page quotes each licence from its canonical file, gives a verdict per model, and takes a clear position on that contradiction instead of dodging it.

This is not legal advice. We are engineers reading licence text, not lawyers; if real money or a client contract depends on the answer, pay someone qualified to read it for your jurisdiction and your business model. Licences also get amended — everything below was read from source on 18 August 2026, and each row links to the file so you can re-check the date stamp yourself.

---

The Short Answer, By What You Are Building

Pick the row that matches you and you are 90% done.

---

The Comparison Table

Here is every model people actually run locally, with the licence exactly as its model card names it. Verdicts are ours; the licence names and quoted text are from the linked source.

Licence fields read from the linked Hugging Face model cards, 18 August 2026. on that date FLUX.2 [dev] showed just over 1.0 million downloads in the previous 30 days and 2.04k likes, and FLUX.1-Fill-dev roughly 345K downloads — which is the scale of the problem: the most-downloaded local image models are the restricted ones. Download counters move daily; check the card for current figures.

---

Outputs vs Weights: Taking A Position

Our position: the images are yours, the pipeline is not. Owning your outputs does not license you to operate the model commercially.

Here is why that is not a dodge. The FLUX [dev] licence contains both of these, verbatim:

"You may only access, use, Distribute, or create Derivatives of the FLUX Model or Derivatives for Non-Commercial Purposes."

"We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs" … "You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein."

And the definition that reconciles them:

Non-Commercial Purpose: "…but only so far as you do not receive any direct or indirect payment arising from the use of the FLUX Model."

Read together, the Outputs clause is about ownership and downstream rights in the image file — BFL is disclaiming any copyright interest in what you generate, which is genuinely useful and better than some competitors. The grant clause is about the act of running the weights. If you receive payment that arises from running the model, that act falls outside the permitted purpose, regardless of who owns the resulting PNG.

So the practical line we would draw:

• Defensible: you generated images as a hobby, and later someone offers to buy a print. The use at generation time was non-commercial; the output is yours to sell.
• Not defensible: a shop, a client deliverable, a paid API, or a subscription product where images are generated because money is coming in. That is payment arising from use of the model.
• Definitely not: fine-tuning [dev] to sell the resulting model, which the licence excludes by name.

If your business depends on the first bullet being read generously, that is a bet on interpretation. Buying the licence or switching the base model both cost less than losing that bet.

---

The FLUX Split, In Detail

The confusing thing about FLUX is that the family contains both the most restricted and one of the most permissive models on this page.

[dev] variants — restricted. FLUX.1 [dev], FLUX.2 [dev] and the tooling variants like FLUX.1-Fill-dev all ship under the non-commercial licence. The escape hatch is written into the text itself: "If you want to use a FLUX Model or a Derivative for any purpose that is not expressly authorized under this License, such as for a commercial activity, you must request a license from Company" — with earlier FLUX.1 [dev] licence text noting such a licence "may be subject to a fee, royalty or other revenue share."

[schnell] — Apache-2.0. The FLUX.1-schnell card is unambiguous: "Released under the apache-2.0 licence, the model can be used for personal, scientific, and commercial purposes." Same family, same architecture lineage, no restriction. It is a distilled, few-step model, so it is faster and lower-fidelity than [dev] — but it is the FLUX you can build a business on without signing anything. Our FLUX local image generation guide covers the practical quality gap between the two.

And the architecture escape route. If what you actually want is FLUX-style output with a clean licence, the more interesting answer is Chroma — an Apache-licensed model built on the FLUX architecture, which exists more or less because of this exact problem. That page is the deep dive; this one is the map.

If you are already running [dev] locally and want to know what you would be giving up by moving, our FLUX.2 local setup guide and the broader best local image models comparison have the quality-and-VRAM side of the trade.

---

Your LoRA Inherits the Base Model

A LoRA trained on FLUX [dev] is a Derivative under that licence, and carries the non-commercial restriction with it — including when you distribute or sell the LoRA file.

The licence defines Derivatives as "any (i) modified version of the FLUX.1 [dev] Model … including customized or fine-tuned versions" or derivative works of it, and the operative grant covers "the FLUX Model or Derivatives" in the same breath. Separately, the Non-Commercial Purpose definition excludes using the model to "train, fine tune, or distill other models for commercial use."

Three consequences people get wrong:

1. Selling the LoRA is distribution of a Derivative. The 200MB file you uploaded is not "your work" in a licensing sense; it is a modification of someone else's restricted weights.
2. Bundling it with a permissive base does not launder it. If the LoRA was trained against [dev], its terms come from [dev].
3. The decision is made at training time and is expensive to reverse. Retraining on Qwen-Image or schnell means redoing the dataset run. Check the base licence before you spend the GPU hours — this is the single most common way people discover the problem too late.

---

The Licences That Are Not Apache

Three non-Apache licences show up constantly, and each permits commercial use with a different string attached.

CreativeML Open RAIL++-M — commercial, with a behaviour clause

The SDXL family licence. SDXL base 1.0 states it as "CreativeML Open RAIL++-M License", and Animagine XL 4.0 adopts Stability's original licence unmodified, summarising it on its own card as "Permitted: Commercial use, modifications, distributions, private use." The "RAIL" part is a use-based restriction: you agree not to use the model for an enumerated list of harmful applications, and you must pass those same restrictions on to anyone you give the model to. There is no revenue cap and no fee. For most commercial art work this is fine; if you are embedding the model in a product, the pass-through obligation is the clause your lawyer will care about.

Stability Community License — commercial under a revenue ceiling

Stable Diffusion 3.5 Large is "free for research, non-commercial, and commercial use for organizations or individuals with less than $1M in total annual revenue." Above that, you need an Enterprise License from Stability AI. Note the threshold is total annual revenue of the organisation, not revenue from the AI images — an agency doing $2M in unrelated work is over the line on day one.

Fair AI Public License 1.0-SD — commercial, but copyleft

The licence on NoobAI-XL 1.1 and much of the anime fine-tune ecosystem. It grants broadly — "Each contributor licenses you to do everything with this software that would otherwise infringe that contributor's copyright in it" — so commercial use is permitted. The obligations are what make it different from Apache:

• Copyleft on derivatives: derived models must be "provided under this license or another license that allows everything this license allows." You cannot fine-tune it and ship the result as proprietary weights.
• A network clause: "If you modify this software and allow users to interact with it through a computer network, you must ensure they have a reasonable way to receive the corresponding source code from you." If you fine-tune it and put it behind a web UI, you owe your users the modified weights.
• A prohibited-uses list, in the RAIL tradition.

Selling images made with it: fine. Building a closed hosted product on a fine-tune of it: read that network clause very carefully first.

---

The Utility-Node Trap

Checking your checkpoint is not checking your pipeline. The upscaler, the background remover, the segmentation model and the face-restoration node each carry their own licence, and at least one of them is probably non-commercial.

The clearest example is briaai/RMBG-2.0, a background remover that sits in a very large number of ComfyUI product-photography graphs. It is CC BY-NC 4.0, and the card is explicit: "Commercial use is subject to a commercial agreement with BRIA." Its own comparison table marks self-hosted weights as ❌ "Requires agreement" for commercial licensing, while the Bria API includes it. Nothing in your pipeline warns you about this — the node loads, the mask is excellent, and the licence question never surfaces.

A five-minute audit that catches almost all of it:

1. List every model file your workflow loads — checkpoint, VAE, LoRAs, ControlNets, upscalers, detectors, segmenters, restorers.
2. Open each one's Hugging Face or Civitai page and read the licence field, not the description.
3. Flag anything containing NC, non-commercial, research only, or requires agreement.
4. For each flag, either find a permissive replacement or price the commercial agreement.

If you run this pipeline for paying clients, our local AI for photographers guide walks through the practical workflow side, and the ComfyUI complete guide covers what each node type is actually doing so you know what to substitute.

---

Buying A FLUX Commercial Licence

If your workflow is genuinely built on [dev] and moving is expensive, BFL sells self-serve licences for the weights — the question is whether the volume caps fit your business.

The licensing page describes the offer as "License FLUX model weights. Fine-tune on your data, self-host on your infrastructure" across four tiers:

Tier names, model coverage and stated limits read from bfl.ai's licensing page, 18 August 2026. Prices are not published there — you purchase or contact sales through their dashboard, so we cannot quote figures we have not seen.

Three things to check before you buy rather than switch. The image cap is per month, not per project — a print-on-demand catalogue refresh can burn 10K generations in a weekend, including the ones you throw away. The domain limit is a real constraint for agencies: one domain on Builder means one storefront, and Platform is also one domain despite the ten-fold image jump. And Builder does not cover [dev] at all — it is a klein tier, so if [dev] is the model you are actually running, Platform is your realistic floor.

---

A Clean Commercial Stack

If you want zero licence ambiguity, this is the whole list. Every model here is Apache-2.0 on its own card, which means no revenue threshold, no domain count, no use-based clause and nothing to sign.

• Text-to-image: Qwen-Image (Apache-2.0) or Z-Image-Turbo (Apache-2.0)
• Image editing: Qwen-Image-Edit-2509 (Apache-2.0)
• Fast / draft passes: FLUX.1 [schnell] (Apache-2.0)
• FLUX-architecture output without the FLUX licence: Chroma (Apache-licensed)
• Anime / illustration: an OpenRAIL++ base such as Animagine XL 4.0, if you accept the use-based restrictions

The honest trade-off: FLUX.2 [dev] is a strong model, and the licence is the reason people put up with it. If your work is not commercial, use it — that is exactly what the licence permits. The moment money enters, the Apache-2.0 set above costs you some quality ceiling and buys you the ability to stop thinking about this entirely. For most Etsy and print-on-demand work, that is a trade worth making. For anything with a client contract behind it, it is not really a trade at all.

---

Limits of This Page

• Not legal advice, and not a substitute for reading the licence yourself. We quote the operative clauses; we do not reproduce entire licences, and clauses we did not quote may matter to your situation.
• Licences change without a version bump. Model cards get amended in place. Every fact here has an 18 August 2026 read date and a link — re-check before you commit a business to it.
• We have not verified every fine-tune. Community checkpoints on Civitai frequently declare terms that conflict with the base model's licence, and a fine-tune cannot grant rights the base did not give it. When the two disagree, the base is the binding one.
• Jurisdiction matters and we ignore it. Copyrightability of AI outputs, and the enforceability of a click-through weights licence, both vary by country. This page is about what the licence says, not what a court where you live will do with it.

---

Sources

• Hugging Face model cards and licence files, read 18 August 2026: FLUX.2-dev, FLUX.1-dev, FLUX.1-Fill-dev, FLUX.1-schnell, Qwen-Image, Qwen-Image-Edit-2509, Z-Image-Turbo, SDXL base 1.0, SD 3.5 Large, Animagine XL 4.0, NoobAI-XL 1.1, RMBG-2.0
• Fair AI Public License 1.0-SD — full licence text
• Black Forest Labs licensing — self-serve commercial tiers for FLUX weights

---

FAQ