Stability AI is a generative artificial intelligence company best known for helping fund and release the Stable Diffusion family of image models. Its principal UK trading entity, Stability AI Ltd, was incorporated on November 4, 2019 and is wholly owned by Stability AI, Inc., a Delaware corporation. The group develops and distributes models for image, video, audio, language, code, and 3D generation through hosted services, self-hosted licenses, custom training, and distribution partners.[1][2]

Stable Diffusion was a collaborative project, not the work of Stability AI alone. Stability AI's 2022 launch announcement credited Patrick Esser of Runway and Robin Rombach of the Machine Vision and Learning group at Ludwig Maximilian University of Munich with leading the work, which built on the CompVis latent diffusion research and received support from Stability AI, LAION, EleutherAI, and other contributors.[3][4][5]

The company calls many of its releases open source, but "open-weight" is a more precise description for much of its current portfolio. Stability AI publishes weights for selected models, while its Community License imposes use conditions and requires a paid Enterprise License for commercial users whose annual revenue exceeds $1 million. Those conditions differ from unrestricted open-source AI licenses.[6][7]

Corporate history and financing

Legal entities and early history

Stability AI Ltd was incorporated in England and Wales in 2019. Its filed accounts identify it as the main trading entity of the Stability AI group and say that its first revenue-generating activities began in August 2022 with the Stable Diffusion series.[1][2]

The history of the US parent is partly contested in litigation. A June 2026 Delaware Court of Chancery opinion recounts that Emad Mostaque, Cyrus Hodes, and a third co-founder incorporated Stability AI, Inc. in October 2020, with Mostaque receiving 70 percent of the equity and the other two receiving 15 percent each. The opinion cautions that this account was drawn from submissions on a discovery dispute and was not a formal finding of fact.[8]

Hodes sold his one million shares to Mostaque in two transactions in 2021 and 2022 for a total of $100. He later alleged fraud, negligent misrepresentation, and breach of fiduciary duty, among other claims. The court dismissed unjust-enrichment and quantum-meruit counts but allowed the three other counts to proceed into discovery. Its June 2026 opinion decided only whether particular communications were protected by spousal privilege, not whether Hodes's underlying allegations were true.[8]

Funding and the 2024 recapitalization

In October 2022, Stability AI announced a $101 million funding round led by Coatue, Lightspeed Venture Partners, and O'Shaughnessy Ventures.[9] In March 2024, Mostaque resigned as chief executive and from the board. Chief operating officer Shan Shan Wong and chief technology officer Christian Laforte became interim co-chief executives.[10]

The company announced another investment round in June 2024 and appointed Prem Akkaraju as chief executive and Sean Parker as executive chairman. The announcement named Greycroft, Coatue, Sound Ventures, Lightspeed, O'Shaughnessy Ventures, Parker, Eric Schmidt, Akkaraju, and Robert Nelsen among the investors but did not disclose the amount.[11] Stability AI Ltd's later accounts say the parent group raised more than $80 million and renegotiated onerous contractual commitments. The UK group recorded a separate $50 million recapitalization from its parent during 2024.[2]

Filed financial results

Stability AI Ltd's accounts for the year ended December 31, 2024 distinguish reported revenue from recurring operations. The group reported $55.0 million in total revenue, but classified $43.6 million as non-underlying. That amount included a one-time $31.0 million dataset-license transaction, $12.2 million from a discontinued managed-services line, and about $0.4 million of intragroup ClipDrop sales. Underlying revenue was $11.4 million.[2]

2024 measure	Amount
Reported revenue	$55.0 million
Underlying revenue	$11.4 million
Loss for the year	$47.5 million
Cash and cash equivalents	$23.9 million
Current liabilities	$81.6 million
Net liabilities	$52.2 million

The accounts compare the full 2024 year with a nine-month 2023 period, so direct year-over-year comparisons require care. They show that the group remained loss-making, with a $47.5 million loss in 2024, down from a $119.5 million loss in the shorter prior period. At year end, it had $23.9 million in cash and cash equivalents and $52.2 million in net liabilities.[2]

The directors prepared the accounts on a going-concern basis, stating that they expected adequate resources for the foreseeable future. They also identified the parent company as the primary source of funds and said it had initiated another funding round when the accounts were signed. The strategic report identified changes in customer demand and continued parent funding as key risks.[2]

Stable Diffusion

Research provenance

Stable Diffusion grew out of research on latent diffusion models by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. Their method first compresses images into a lower-dimensional latent representation, then trains a diffusion process in that space. This reduces the computing cost of image synthesis compared with running the entire denoising process at full pixel resolution.[4]

The original Stable Diffusion repository describes the project as a collaboration among CompVis, Stability AI, and Runway. It says Stability AI donated computing resources and LAION provided support, and that the model was trained on subsets of LAION-5B.[5] Stability AI's own August 2022 announcement similarly credited Esser and Rombach as project leads and named CompVis, Runway, LAION, EleutherAI, and the company's generative AI team as contributors.[3]

Public release and licensing

Stability AI announced the public release on August 22, 2022. Model weights and code were made downloadable under the CreativeML OpenRAIL-M license, which permitted commercial and non-commercial use but included use-based restrictions. The original repository also warned that the weights had known limitations and biases and recommended additional safety mechanisms for products and services.[5][12]

The reference implementation included a safety checker in its default generation pipeline.[5] Because the weights and code could be run and modified outside Stability AI's infrastructure, that component was not equivalent to a mandatory platform-wide control. The release therefore combined broad access and local adaptability with fewer centralized enforcement mechanisms than a hosted-only service.

Later image-model generations

Stable Diffusion XL, or SDXL, expanded the architecture with a UNet about three times larger than earlier Stable Diffusion versions, a second text encoder, additional conditioning, and an optional refinement stage. The accompanying paper described a two-stage base-and-refiner pipeline and evaluated the model primarily through human preference studies.[13]

The Stable Diffusion 3 research introduced a multimodal diffusion transformer architecture and a rectified-flow formulation. The paper studied scaling up to an 8-billion-parameter model and focused on prompt following, typography, and text-to-image quality.[14]

Stability AI released Stable Diffusion 3.5 Large and Large Turbo in October 2024, followed by Medium later that month. The company described Large as an 8.1-billion-parameter base model, Large Turbo as a four-step distilled version, and Medium as a 2.5-billion-parameter model intended for consumer hardware.[15] These models use the Stability AI Community License rather than the original CreativeML OpenRAIL-M terms.[15][7]

Model portfolio

Stability AI's May 2026 Core Models list groups its licensed portfolio into media, language and code, 3D, and audio. The list is a current licensing record rather than a permanent product catalog, and the company expressly reserves the right to modify it.[6]

Area	Representative model families	Documented role
Image	Stable Diffusion 3.5, Stable Diffusion 3, SDXL Turbo	Text-to-image generation and image editing
Video	Stable Video Diffusion	Image-to-video generation in 14-frame and 25-frame variants
Language and code	Stable LM 2, Stable LM Zephyr, Stable Code	Text and code generation
3D	Stable Fast 3D, Stable Video 3D, Stable Point Aware 3D	Reconstruction or generation of 3D assets
Audio	Stable Audio Open, Stable Audio 3	Music and sound-effect generation

Video, audio, language, and 3D

The Stable Video Diffusion research adapted latent video diffusion models to text-to-video and image-to-video generation. Its released image-to-video variants generated 14 or 25 frames, and the work also explored multi-view generation for 3D applications.[16]

Stable Audio Open, released in 2024, is a text-to-audio model documented as having been trained only on Creative Commons licensed audio. Its paper describes generation of stereo audio at 44.1 kHz for clips up to 47 seconds and presents it mainly as a model for sound effects and short production elements rather than complete songs.[17]

The May 2026 Stable Audio 3 technical report describes small, medium, and large latent diffusion models for variable-length music and sound-effect generation, editing, inpainting, and continuation. It says the models were trained on licensed and Creative Commons data. The small and medium weights were released, while the product announcement made the Large model available through the API and enterprise self-hosting. Medium and Large support outputs of up to 6 minutes 20 seconds.[18][19]

Stable LM began in April 2023 with 3-billion and 7-billion parameter base models. The base releases used a CC BY-SA 4.0 license, while the first instruction-tuned research versions used a non-commercial license, an important distinction within the family.[20] The company's later Core Models list includes Stable LM 2 and Stable Code variants under its current licensing program.[6]

Stable Fast 3D reconstructs a textured, UV-unwrapped mesh from a single image. Its paper reports approximately 0.5 seconds of generation time on an Nvidia H100 GPU and describes direct mesh prediction, material estimation, and fast UV unwrapping.[21]

Access and licensing

The filed accounts describe four principal routes to market: models hosted by Stability AI, self-hosted annual licenses that may include custom training and support, distribution partners operating under revenue-share or listing arrangements, and applications sold by subscription. Hosted access can be sold through web interfaces, APIs, credits, subscriptions, or enterprise seats.[2]

For current Core Models, the Community License permits research, non-commercial use, and commercial use by people or organizations with less than $1 million in annual revenue. Commercial users above that threshold must obtain an Enterprise License. The license also incorporates an Acceptable Use Policy and restricts using Core Models or their outputs to create a competing foundational model, although it permits fine-tunes and LoRAs.[7][22]

These terms allow inspection and local deployment of selected weights but are not the same as an unconditional open-source grant. Individual releases can also carry different licenses, so the terms for a specific model and version must be checked rather than inferred from the Stability AI brand.[6][7]

Business and partnerships

The group's reported business model combines hosted inference, self-hosted model licenses, custom model training, applications, and partner distribution.[2] Amazon Web Services added Stable Image Ultra, Stable Diffusion 3 Large, and Stable Image Core to Amazon Bedrock in September 2024, giving customers a managed route to those models through AWS.[23]

Stability AI and WPP announced an investment and partnership in March 2025. Their announcement described joint research and development and integration of Stability AI workflows into WPP Open. Stability AI Ltd's accounts independently list WPP among the group's strategic partnerships.[2][24]

Arm reported in March 2025 that the companies had optimized a distilled Stable Audio Open model to run locally on Arm-based smartphones. Arm said the collaboration used KleidiAI, XNNPack, and ExecuTorch and reduced generation time from minutes to seconds in its demonstration.[25]

Electronic Arts announced a partnership with Stability AI in October 2025 to co-develop generative tools and workflows for game artists and developers.[26] Universal Music Group announced a separate alliance that month to develop professional music-creation tools, with an emphasis on artist involvement and licensed music data.[27] These announcements describe intended collaborations; they are not evidence by themselves of revenue or broad adoption.

In April 2026, Stability AI launched Brand Studio, a managed creative-production platform for professional and enterprise teams. The product combines custom brand models, campaign references, editing tools, and model routing that may use both Stability AI and third-party models.[28] The launch marked a further shift toward managed applications and enterprise workflows alongside downloadable model weights.

Leadership and governance

Mostaque resigned as chief executive and from the board in March 2024. The company said he left to pursue decentralized AI, while Wong and Laforte served as interim co-chief executives.[10] Prem Akkaraju became chief executive in June 2024, and Sean Parker joined as executive chairman.[11]

As of the July 2026 research cutoff, Stability AI's board page listed Sean Parker, Prem Akkaraju, James Cameron, Dana Settle, and Thomas Laffont.[29] The company's current leadership structure therefore differs substantially from the founder-led organization that released the first Stable Diffusion weights.

Data, safety, and litigation

Getty Images litigation

Getty Images brought proceedings in both the United Kingdom and the United States over Stable Diffusion. The outcomes must be separated because the cases involve different law, allegations, and procedural stages.

In the UK case, Getty abandoned its training-and-development claim after acknowledging that there was no evidence the relevant training occurred in the United Kingdom. It also abandoned its output-copyright claim after Stability AI blocked the prompts at issue, and its database-right claim could no longer proceed.[30]

The November 2025 High Court judgment rejected Getty's secondary copyright theory. The court held that the final model did not store or reproduce the copyrighted works and therefore was not an "infringing copy" for the relevant secondary-infringement provisions. Getty did succeed on limited trademark claims involving a few historic outputs from Stable Diffusion 1.x and 2.x that resembled Getty or iStock watermarks. The judge described those findings as "historic and extremely limited in scope" and found no corresponding UK-user evidence for SDXL or version 1.6.[30]

That judgment did not decide whether training Stable Diffusion in another country infringed copyright there. The court explicitly confined itself to the diminished UK case before it, so it should not be read as a general ruling that Stable Diffusion training or outputs are lawful everywhere.[30]

Getty's US litigation did not end when it left Delaware. Getty voluntarily dismissed the Delaware action without prejudice on August 14, 2025 and filed a new complaint in the Northern District of California the same day.[31] In April 2026, the California court considered a motion directed at counts II through VII. It dismissed the false copyright-management-information count without prejudice and allowed Getty to amend it, but declined to dismiss the challenged federal trademark, false-designation, dilution, and California unfair-competition claims. The copyright-infringement count was not part of that motion.[31]

Artist class action and founder dispute

The artist lawsuit led by Sarah Andersen also remained active. A June 17, 2026 discovery order required the plaintiffs to provide some additional financial information and discussed the unsettled application of fair use to generative AI. The order resolved a discovery dispute, not the merits of the copyright claims.[32]

The Hodes founder litigation likewise remained in discovery as of June 2026. As noted above, the surviving fraud, negligent-misrepresentation, and fiduciary-duty claims had not been adjudicated on the merits.[8]

LAION-5B and child-safety concerns

The first Stable Diffusion release was trained on subsets of LAION-5B, a dataset maintained by the independent nonprofit LAION rather than by Stability AI.[3][5] LAION-5B distributed metadata, including image URLs, text descriptions, and embeddings, rather than a bundle of the referenced image files.[33]

A 2023 Stanford Internet Observatory study identified 3,226 LAION-5B entries as suspected child sexual abuse material, or CSAM. Of those, 1,008 were manually validated by the Canadian Centre for Child Protection as CSAM or likely CSAM. The report stressed that many URLs were dead or could not be conclusively classified and that the dataset entries were links and metadata, not stored copies of the images.[33]

LAION took LAION-5B and its derivatives offline in December 2023. It released Re-LAION-5B in August 2024 after working with the Internet Watch Foundation, the Canadian Centre for Child Protection, and the Stanford researchers. LAION said it removed 2,236 links that matched partner-provided image or URL hashes, including the 1,008 validated links from the Stanford report. Its broader filtering for the research release removed about 1.121 percent of the original samples, or approximately 65 million entries from the 5.8-billion-entry source.[34]

The revised metadata did not retroactively change model weights that had already been trained and released. It did, however, provide removal lists that dataset users could apply to derivatives of LAION-5B.[34]

Company-stated safeguards and their limits

Stability AI says that, for model versions it develops exclusively, it filters training data to remove unsafe images. It also says its hosted applications and APIs filter prompts and outputs, use red teaming and safety fine-tuning, compare uploads against known-CSAM hashes, and add C2PA provenance metadata to API-generated media.[35][36]

The company-wide Acceptable Use Policy applies contractually to hosted access, third-party platforms, and self-hosted code or weights.[22] Technical enforcement differs by deployment route, however. The prompt and output filters described by Stability AI operate in its hosted applications and APIs; independently run weights do not pass through that hosted moderation layer. This distinction is one of the central tradeoffs of open-weight distribution.[35][22]

In February 2026, Stability AI announced that it had joined the Tech Coalition after participating in the coalition's Pathways program since 2025.[37] Membership is evidence of industry participation, while the effectiveness of the company's safeguards still depends on implementation, coverage, and the deployment environment.

Developments through July 2026

By the July 28, 2026 cutoff, Stability AI was pursuing two connected strategies. It continued to publish selected weights, including the small and medium Stable Audio 3 models, while selling enterprise licenses, hosted APIs, custom training, and managed applications.[2][7][19]

Its 2026 releases and announcements included Tech Coalition membership in February, Brand Studio in April, and Stable Audio 3 in May.[37][19][28] Together with the 2025 partnerships in advertising, games, mobile hardware, and music, these developments indicate a stronger focus on professional creative workflows and licensing. They do not erase the company's continuing dependence on parent financing or the unresolved US litigation described in its filed accounts and court records.[2][31][32]

Editing: Stability AI