Free vs. Paid: What You Are Actually Paying For
"It's the same Stable Diffusion, so why do some sites charge?"
This trips up a lot of people. Search for Stable Diffusion and you get pages of paid web services. Then you find the same name on GitHub, free to download. It looks suspicious, but both are legitimate, and the GitHub route is free and legal. The real difference is whose computer runs the model.
- Local (free): Stable Diffusion models are released as open weights. You download a front end such as ComfyUI or AUTOMATIC1111's WebUI from GitHub, grab a model file from Hugging Face or Civitai, and generate images on your own graphics card. Beyond electricity, it costs nothing.
- Web (paid): Stability AI's own API, plus many third-party sites that host Stable Diffusion on their servers. You are paying to rent someone else's GPU, not buying the right to use Stable Diffusion itself.
In other words, paid web services sell convenience: no install, no high-end GPU, just a few clicks. Stability AI's API bills in credits. As of November 6, 2026, one credit costs $0.01, and one image takes 3 credits with Stable Image Core, 6.5 with SD 3.5 Large, and 8 with Stable Image Ultra. That works out to roughly 3 to 8 cents per image. Third-party sites set their own prices. If you own a capable graphics card, you can run the same open models locally with no per-image fee. Open release was a deliberate choice by the developer, so running the model for free is not piracy.
Stable Diffusion licensing comes in three layers: (1) the license of the program you run (ComfyUI, WebUI, and so on), (2) the license of the model file, or checkpoint, and (3) the rights in the images you generate. Each layer has its own rules. Most of the confusion comes from lumping the three together.
Program Licenses: Why You Rarely Need to Worry
ComfyUI is GPL-3.0, AUTOMATIC1111 WebUI is AGPL-3.0
The front ends you download from GitHub each carry an open-source license. AUTOMATIC1111's Stable Diffusion WebUI uses AGPL-3.0, and ComfyUI uses GPL-3.0 (its GitHub repository is now served under the Comfy-Org organization). If you have heard that "the GPL limits commercial use," this is where people get nervous. For someone who simply uses the tool, it rarely matters.
GPL-family licenses attach conditions to modifying and distributing the program, and the AGPL adds a rule for network use. If you change ComfyUI's code and ship it as your own product, you have to share the source under the GPL. If you modify the WebUI and let people use it over a network, the AGPL requires you to offer those users the modified source. Neither of those describes what most people do.
- Installing it on your PC and making images: Allowed, for personal or work use. The GNU GPL FAQ also says you are "free to make modifications and use them privately, without ever releasing them."
- Selling the images you make: Not a program-license question. The same FAQ states that "the output of a program is not, in general, covered by the copyright on the code of the program." A novel you write in a word processor does not fall under the word processor's license.
- Where to be careful: Selling a modified version of the program, or running a modified version as a service for other people. Then you need to work through the GPL or AGPL terms, including the source-code obligations.
Whether you are making blog thumbnails or mockups for a client, the program license is not the issue as long as you use the software as a tool. It starts to matter when you resell the tool itself or turn it into a service.
The Model License Is the One That Matters
Each model version has its own rules
This is the layer you actually need to check. Licenses attach to each model file (checkpoint), not to the program. Two models can both be called "Stable Diffusion" and still come with different terms. Starting with the 2024 releases, Stability AI also added a revenue threshold.
| Model | License | Commercial use |
|---|---|---|
| SD 1.5 | CreativeML OpenRAIL-M | Allowed, no revenue cap (use-based restrictions apply) |
| SDXL 1.0 (base) | CreativeML Open RAIL++-M | Allowed, no revenue cap (use-based restrictions apply) |
| SD 3.5 (Large, Large Turbo, Medium) | Stability AI Community License | Free under $1M in annual revenue; Enterprise License above that |
| FLUX.1 [schnell] | Apache 2.0 | Allowed for personal, scientific, and commercial use |
| FLUX.1 [dev] | FLUX.1 [dev] Non-Commercial License | Model: non-commercial only. Outputs: commercial use allowed, with exceptions in the license |
OpenRAIL: commercial use is fine, misuse is not
SD 1.5 uses the CreativeML Open RAIL-M license (dated August 22, 2022), and SDXL 1.0 uses the CreativeML Open RAIL++-M License (dated July 26, 2023). Both grant a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license." There is no revenue cap and no royalty, whether your use is commercial or not. Both also say: "Licensor claims no rights in the Output You generate using the Model."
What they add are use-based restrictions, listed in Attachment A. The SDXL license, for example, bars unlawful use, harming minors, spreading harmful false information, defamation or harassment, and discrimination, among other things. Normal creative or business work does not run into these. If you redistribute the model or a fine-tune of it, the license requires you to pass the same use-based restrictions on to the people you share it with.
One naming trap for anyone searching "SDXL license": SDXL Turbo is not under OpenRAIL++-M. Stability AI's license page lists SDXL Turbo among the models covered by the Community License, alongside the Stable Diffusion 3.5 suite.
SD 3.5: the $1 million revenue line
SD 3.5 Large and Large Turbo were announced on October 22, 2024, and SD 3.5 Medium followed on October 29. They ship under the Stability AI Community License (last updated July 5, 2024). The model card puts it plainly: free for research, non-commercial, and commercial use "for organizations or individuals with less than $1M" in total annual revenue. Above that, Stability AI asks you to contact it for an Enterprise License. Most solo bloggers, freelancers, and small studios sit well inside the free tier.
A few details are easy to miss:
- Registration: For commercial purposes, the agreement says "You must register with Stability AI."
- Outputs: "You own any outputs generated from the Models or Derivative Works to the extent permitted by applicable law." The revenue tier does not change this.
- Acceptable use: You must follow Stability AI's Acceptable Use Policy, which the agreement incorporates by reference.
- Attribution: If you distribute the models or derivative works, the agreement requires a prominent "Powered by Stability AI" notice on a related website, user interface, blog post, about page, or product documentation.
FLUX.1: [schnell] vs. [dev]
FLUX.1 [schnell] uses Apache 2.0, and its model card says it "can be used for personal, scientific, and commercial purposes." FLUX.1 [dev] works differently. Its FLUX.1 [dev] Non-Commercial License (v1.1.1) limits the model itself to non-commercial purposes. Revenue-generating activity and training models for commercial use fall outside that definition.
Outputs are treated separately. The license says Black Forest Labs claims "no ownership rights in and to the Outputs," and "You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein." One of those prohibitions: you may not use outputs to train, fine-tune, or distill a model that competes with FLUX.1 [dev]. For commercial use of the model itself, the license points you to bfl.ai.
Community models: check each one
Fine-tuned checkpoints and LoRAs on Civitai (RealVisXL, Pony, and many others) build on a base model and its license, and each creator can add conditions on top. Civitai lets creators set permissions per model, and the model page shows them as icons: green means allowed, red means restricted, and hovering over an icon explains it. These are the options a creator can set:
- General: "Use without crediting me," "Share merges of this model," and "Use different permissions on merges."
- Commercial use: "Sell generated images," "Use on Civitai generation service," "Use on other generation services," "Sell this model," and "Sell merges using this model."
Get in the habit of checking those icons before you download.
When you download model files, pick the .safetensors format over .ckpt. Older .ckpt checkpoints use Python's pickle format, and Hugging Face's security docs warn of "dangerous arbitrary code execution attacks that can be perpetrated when you load a pickle file." Safetensors was built to store model weights safely, without pickle. This is basic security hygiene, separate from licensing.
Can You Use the Images Commercially?
Using an image and owning a copyright are two different questions
Short answer: commercial use of generated images is fine as long as you follow the model license above. Blog posts, YouTube thumbnails, book covers, product designs, and client work all qualify. The OpenRAIL licenses say the licensor claims no rights in your output. The Stability AI Community License says you own your outputs to the extent the law allows. The FLUX.1 [dev] license says Black Forest Labs claims no ownership of outputs.
Just don't confuse "I'm allowed to use it" with "it's protected as my work." In the United States, the Copyright Office's report on copyrightability, released January 29, 2025, concludes that AI outputs are protected only where a human author has determined sufficient expressive elements, for example through creative arrangements or modifications of the output, "but not the mere provision of prompts." So an image the AI produced largely on its own may be hard to stop others from copying. Where you add real human authorship, such as compositing, editing, or arranging, that human-made part may be protectable.
Steer clear of these: the real risks sit outside the license
- Real people: If you generate a celebrity's face and use it commercially, the problem is not the license. It is likeness and right-of-publicity law.
- Existing characters and logos: Selling merch that looks like a famous character can infringe copyright or trademark. "I made it with AI" is not a defense.
- LoRAs trained on one artist's style: The law here is unsettled, but using a LoRA trained on a living artist's style for commercial work carries real dispute and reputation risk.
- Platform rules: Print shops, stock-image sites, and contests each have their own AI policies. Check the terms of every channel where you sell.
This article summarizes publicly available license documents for general information. It is not legal advice. If you are applying this to a sizable business or a contract, have a lawyer review it.
Q&A: Common Situations
Q1. Can I use the images for blog thumbnails or YouTube backgrounds if the site earns ad revenue?
Yes. SDXL 1.0 has no revenue cap even with ad income, and SD 3.5 stays free while your total annual revenue is under $1M (register with Stability AI for commercial use). Few individual creators come close to that line.
Q2. Can I sell merch (T-shirts, stickers) with images I generated?
Under the model licenses, yes. If you used a Civitai community model, check the "Sell generated images" icon on its page. Avoid images that resemble existing characters or real people.
Q3. Can I use it for work (mockups, slide decks, marketing images)?
SD 1.5 and SDXL 1.0 work regardless of company size. SD 3.5 needs an Enterprise License once the company's annual revenue tops $1M. A larger company can stick with SDXL 1.0 or FLUX.1 [schnell], or sign the license.
Q4. Can I mix AI images into client deliverables?
The licenses allow it. The practical issue is the contract. AI-generated material may have weak copyright protection, so think about how your warranty or IP clauses handle that. Telling the client you used AI is the safer path.
Q5. What if I want to add image generation to my own website as a service?
At that point you are a service provider, not just a tool user. The program's AGPL or GPL obligations, whether the model allows hosted use, and the SD 3.5 revenue terms all come into play, so review all three layers. On Civitai, look at "Use on other generation services." The simplest route is often a paid API such as Stability AI's, which can end up cheaper.
Q6. So when does a paid web version make sense?
When you have no capable GPU (say, a laptop with integrated graphics), need only a few images now and then, or don't want to spend time installing and maintaining software. If you generate dozens of images every day, running locally is far more economical. The difference is convenience, not capability.
A Checklist for Staying Safe
To sum up: Stable Diffusion from GitHub is free and legal for personal and small-business commercial use, as long as you follow each model's license. Build these five habits:
- 1. Check the license on the model page first: On Hugging Face, look at the License field. On Civitai, look at the permission icons. It takes ten seconds before you download.
- 2. Use SDXL 1.0 as your default for commercial projects: It has no revenue terms. For SD 3.5, confirm you are under $1M in annual revenue and register for commercial use. For FLUX.1, know whether you have [schnell] (Apache 2.0) or [dev] (non-commercial model).
- 3. Download .safetensors only: Skip .ckpt files from unknown sources.
- 4. Keep real people, existing characters, and living artists' styles out of commercial work: Much of the real risk sits there, not in the license.
- 5. Record which model you used: For images you sell or deliver, note the model and the date. If a license question comes up later, you can answer right away.
Wrapping up
A paid web version does not make the free version a bootleg. Stable Diffusion was released as open weights from the start, and paid services sell GPU time and convenience. As long as you use the software as a tool, the program license is not a concern. Check the model license version by version, and individuals and small businesses can use these models commercially for free. For setup, see the Stable Diffusion local install guide. For the next step in workflows, there is a ComfyUI beginner's guide (in Korean).
FAQ
Is SDXL free for commercial use?
SDXL 1.0 base is released under the CreativeML Open RAIL++-M License, which has no revenue cap and no royalty. You still have to follow its use-based restrictions. SDXL Turbo is a separate model covered by the Stability AI Community License.
What does the $1 million limit in the Stability AI Community License mean?
Individuals and organizations with less than $1M in total annual revenue can use covered models such as SD 3.5 for free, including for commercial purposes. Above $1M, Stability AI asks you to obtain an Enterprise License. Commercial users also need to register with Stability AI.
What are the model licensing options on Civitai?
Creators can allow or restrict "Use without crediting me," "Share merges of this model," "Use different permissions on merges," "Sell generated images," "Use on Civitai generation service," "Use on other generation services," "Sell this model," and "Sell merges using this model." The model page shows each choice as a green or red icon.
Can I sell images made with FLUX.1 [dev]?
The FLUX.1 [dev] Non-Commercial License says you may use Output for any purpose, including commercial purposes, except as the license prohibits, such as using outputs to train a competing model. Using the model itself commercially requires a separate license from Black Forest Labs.
Are AI-generated images protected by copyright?
In the US, the Copyright Office's January 2025 report says AI output is protected only where a human author has determined sufficient expressive elements, and prompts alone are not enough. Human edits, arrangements, or other contributions may be protectable.
References
- Stability AI, Stability AI Community License Agreement (last updated July 5, 2024), stability.ai/community-license-agreement (checked November 6, 2026)
- Stability AI, License overview (Community and Enterprise), stability.ai/license (checked November 6, 2026)
- Stability AI, "Introducing Stable Diffusion 3.5," stability.ai/news-updates/introducing-stable-diffusion-3-5 (checked November 6, 2026)
- Hugging Face, stabilityai/stable-diffusion-3.5-large model card, huggingface.co/stabilityai/stable-diffusion-3.5-large (checked November 6, 2026)
- Hugging Face, stabilityai/stable-diffusion-xl-base-1.0 model card and LICENSE.md (CreativeML Open RAIL++-M), huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 (checked November 6, 2026)
- Hugging Face, stable-diffusion-v1-5/stable-diffusion-v1-5 model card, huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 (checked November 6, 2026)
- CompVis, CreativeML Open RAIL-M license text, github.com/CompVis/stable-diffusion (LICENSE) (checked November 6, 2026)
- Black Forest Labs, FLUX.1 [dev] Non-Commercial License v1.1.1, huggingface.co/black-forest-labs/FLUX.1-dev (checked November 6, 2026)
- Hugging Face, black-forest-labs/FLUX.1-schnell model card, huggingface.co/black-forest-labs/FLUX.1-schnell (checked November 6, 2026)
- ComfyUI repository (GPL-3.0), github.com/comfyanonymous/ComfyUI (checked November 6, 2026)
- AUTOMATIC1111, stable-diffusion-webui repository (AGPL-3.0), github.com/AUTOMATIC1111/stable-diffusion-webui (checked November 6, 2026)
- GNU Project, Frequently Asked Questions about the GNU Licenses, gnu.org/licenses/gpl-faq.html (checked November 6, 2026)
- Civitai Education, "Guide to Civitai Model Licensing Options," education.civitai.com/guide-to-licensing-options-on-civitai (checked November 6, 2026)
- Stability AI, API pricing, docs.stability.ai/pricing (checked November 6, 2026)
- U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 29, 2025), copyright.gov/newsnet/2025/1060.html (checked November 6, 2026)
- Hugging Face, Pickle Scanning (Hub security docs), huggingface.co/docs/hub/security-pickle; Safetensors documentation, huggingface.co/docs/safetensors (checked November 6, 2026)
