In brief
- Stable Diffusion users pushed back last month after Stability AI launched restrictive new licensing terms.
- The firm announced late Friday that it has relaxed its conditions, to mixed reactions from users.
- Users still cannot create a new foundational model by training it on SD3-generated work.
Stability AI unveiled a revamped Community License for its Stable Diffusion 3 (SD3) model, aiming to quell the firestorm of controversy that erupted following the initial release. The company's move comes on the heels of a ban by CivitAI, a major community hub, which had barred all SD3-related content due to licensing concerns.
“We acknowledge that our latest release, SD3 Medium, didn’t meet our community’s high expectations,” Stability said in an announcement late Friday. “We heard you and have made improvements to address your concerns and to continue to support the open-source community."
Under the new terms, Stability AI grants free use of SD3 for research, non-commercial, and limited commercial purposes. The license also allows individuals and businesses with annual revenues under $1 million to use the model without charge. Those exceeding this threshold must obtain a paid enterprise license.
In an interview with Stability, the company confirmed that it is OK to create custom SD3 models and improve on top of the base SD3. However, it’s forbidden to develop a new foundational model using images generated with SD3 as part of its training dataset—that is, training a Stable Diffusion competitor using material from the original model.
“Derivative products include any output derived from Stability AI's Foundational models, such as fine-tuned models or other creative outputs,” a spokesperson from Stability AI told Decrypt. “Examples of derivative works include SD3 fine-tunes, LoRA fine-tunes, adapters etc. and these can also be trained with SD3 output images.”
The license also says that “you are the owner of derivative works you create, subject to Stability AI’s ownership of the Stability AI materials and any derivative works made by or for Stability AI.” In other words, as long as those boundaries are respected, fine-tuning and profiting from it should not be against the terms and conditions.
“To safeguard our IP, it is not permitted to train new foundational AI models using SD3 outputs as training data, and all activity must adhere to our acceptable use policies,” the company spokesperson told Decrypt.
Stability AI claims that the updated license recommits to open-source principles, emphasizing creator freedom, research alignment, and community focus. The company pledged to remove restrictions on the number of generated media files, and promised that users below the revenue threshold won't need to delete their outputs or fine-tuned models.
The icing on the cake came with the company promising to update its model to address problems that users found with the first version, namely the disastrous depiction of human anatomy that famously led it to create monsters when asked to generate people lying on the floor.
“We are constantly researching improvements on different technical fronts and we will release these details along with the next model release,” the spokesperson told Decrypt.
A “one star” review
Despite Stability AI's efforts to smooth things over, not everyone in the AI community is convinced. Kent Keirsey, CEO of Invoke AI, expressed skepticism about the new license. Keirsey argued that the revamped terms fail to resolve fundamental issues and introduce new complications.
“I'm going to have to say big old thumb comes down to SD3 new license,” he said during a YouTube analysis of the new terms and conditions. “I know that it was advertised as solving the problems, [but] it doesn't really solve the problems that I saw for people who are actually concerned about building a business around this stuff.”
Keirsey pointed out several problematic aspects of the license, including the fact that Stability AI can terminate the agreement at will. Although this is normal for many software licenses, it is not a part of the most important open-source licenses such as the GPL or Apache licenses. In fact, the license for the previous SDXL model was irrevocable.
“I'm going to give this one star," he concluded. "This is a one star license, maybe one-and-a-half stars [because] it's better than the original."
Edited by Andrew Hayward