Stability AI
Last reviewed
Jun 2, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v9 · 6,651 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v9 · 6,651 words
Add missing citations, update stale details, or suggest a clearer explanation.
Stability AI is a British artificial intelligence company best known as the creator of Stable Diffusion, one of the most widely used open-source text-to-image models. Founded in 2020 by Emad Mostaque and Cyrus Hodes, the company is headquartered in London and has built a portfolio of generative AI products spanning image, audio, video, language, code, and 3D generation. Stability AI's stated mission is "AI by the people, for the people," reflecting its emphasis on open-source development and broad accessibility.
The company rose to global prominence with the August 2022 release of Stable Diffusion and quickly attracted venture capital investment, reaching a $1 billion valuation. However, Stability AI's trajectory has included periods of financial strain, leadership upheaval, copyright litigation, and intense competition. After founder Emad Mostaque departed in March 2024, the company restructured under new CEO Prem Akkaraju and pivoted toward enterprise customers, reporting triple-digit revenue growth by late 2024 and a landmark UK court win against Getty Images in November 2025.[1][21]
Stability AI Ltd was incorporated on November 4, 2019, in England and Wales under company number 12295325. The company was co-founded in late 2020 by Emad Mostaque and Cyrus Hodes. Its registered office is at Fora-United House, 9 Pembridge Road, Notting Hill Gate, London, with its main headquarters at 88 Notting Hill Gate.
Emad Mostaque is a British-Bangladeshi businessman and mathematician. He graduated from the University of Oxford in 2005 with a degree in mathematics and computer science. After university, he spent roughly 13 years working as an analyst at various hedge funds in the United Kingdom. His interest in AI deepened after his son was diagnosed with autism, prompting him to explore how AI could be used to review and synthesize existing medical research.
Before founding Stability AI, Mostaque launched Symmitree in 2019, a venture aimed at reducing the cost of technology for people in developing countries. That project did not succeed. During the early stages of the COVID-19 pandemic in 2020, Mostaque and Hodes led an effort called the Collective and Augmented Intelligence Against COVID-19 (CAIAC) project, which sought to use big data and AI to assist governments with pandemic-related decision-making. This collaboration ultimately led to the creation of Stability AI.
Cyrus Hodes, a French-American AI policy expert who had worked at the OECD and advised governments on AI governance, served as co-founder. According to later court filings, Hodes worked full-time for 18 months building the company, including fundraising and marketing. In 2021 and 2022, Hodes sold his 15% stake in Stability AI for just $100. In July 2023, Hodes filed a civil lawsuit against Mostaque, alleging he had been defrauded into selling his stake at that token price, a stake later estimated to be worth over $500 million.[24]
Stability AI's early operations were self-funded by Mostaque. The company's breakout moment came in October 2022, when it announced a $101 million seed funding round led by Coatue Management and Lightspeed Venture Partners, with participation from O'Shaughnessy Ventures. The round closed on September 19, 2022, valuing the company at approximately $1 billion and making it a unicorn at the seed stage.[2][3][4]
Mostaque stated at the time that the new capital would support deploying custom versions of Stable Diffusion at larger scale, investing in additional supercomputing power, and expanding the team.
In June 2024, following Mostaque's departure and a period of acute financial difficulty, Stability AI announced a major recapitalization led by entrepreneur Sean Parker and venture firm Greycroft. The deal combined three elements:[29][30]
The transaction was structured as an effective takeover of the company by the new investor consortium. It coincided with the appointment of Prem Akkaraju as CEO and Sean Parker as Executive Chairman.
In March 2025, advertising conglomerate WPP announced a strategic investment in Stability AI as part of a broader partnership. While the investment amount was not publicly disclosed, sources indicated it formed part of an extension to the 2024 funding round. Total funding raised by Stability AI through early 2025 exceeded $225 million.[20]
Stability AI's defining product is Stable Diffusion, a latent diffusion model for text-to-image generation. The foundational research was conducted by Robin Rombach, Patrick Esser, and their colleagues at the CompVis Group at Ludwig Maximilian University of Munich (LMU Munich), under the supervision of Professor Bjorn Ommer. Stability AI provided the computing resources and funding needed to train the full-scale model on the LAION-5B dataset, a massive open dataset of 5.85 billion image-text pairs curated by the LAION nonprofit.
Stable Diffusion 1.4 was publicly released on August 22, 2022, with its model weights made freely downloadable. This was unprecedented for a model of this capability. With its 860 million parameter UNet and 123 million parameter text encoder, the model was lightweight enough to run on consumer GPUs with as little as 10 GB of VRAM. The open release stood in stark contrast to proprietary competitors like DALL-E 2 from OpenAI and Midjourney, which were only available through closed platforms.
The impact was immediate. Stable Diffusion was downloaded more than 200,000 times shortly after release. DreamStudio, Stability AI's own hosted interface for the model, quickly surpassed one million registered users across more than 50 countries, with users generating over 170 million images.
Stability AI has released a broad range of generative AI models across multiple modalities. The following table summarizes the company's major product releases.
| Product | Category | Initial Release | Description |
|---|---|---|---|
| Stable Diffusion 1.4 | Image generation | August 2022 | First public release of the latent diffusion text-to-image model |
| Stable Diffusion 1.5 | Image generation | October 2022 | Improved version with refined classifier-free guidance |
| Stable Diffusion 2.0 | Image generation | November 2022 | New text encoder (OpenCLIP), higher resolution support |
| Stable Diffusion 2.1 | Image generation | December 2022 | Refinements to version 2.0 based on community feedback |
| DreamStudio | Platform | August 2022 | Hosted web interface and API for Stable Diffusion |
| StableLM | Large language model | April 2023 | Open-source language models with 3B and 7B parameters |
| Stable Diffusion XL (SDXL) 1.0 | Image generation | July 2023 | Larger architecture with base and refiner models at 1024x1024 resolution |
| Stable Code Alpha | Code generation | August 2023 | Code completion model for 18 programming languages |
| Stable Audio 1.0 | Audio generation | September 2023 | Text-to-audio model producing 44.1 kHz music and sound effects |
| Stable Video Diffusion (SVD) | Video generation | November 2023 | Image-to-video model generating 14 or 25 frames at up to 576x1024 |
| SDXL Turbo | Image generation | November 2023 | Single-step image generation using Adversarial Diffusion Distillation |
| Stable Code 3B | Code generation | January 2024 | 3B parameter code completion model with Fill-in-the-Middle support |
| Stable Cascade | Image generation | February 2024 | Three-stage Wurstchen architecture, 16x cheaper to train than SD3 baseline |
| TripoSR | 3D generation | March 2024 | Image-to-3D model generating textured meshes in under one second |
| Stable Audio 2.0 | Audio generation | April 2024 | Full-length tracks up to 3 minutes with audio-to-audio generation |
| Stable Diffusion 3 Medium | Image generation | June 2024 | New MMDiT Diffusion Transformer architecture, improved text rendering |
| Stable Audio Open 1.0 | Audio generation | June 2024 | Open-weights model trained on CC-licensed audio |
| Stable Video 4D | 4D generation | July 2024 | Video-to-multi-view-video synthesis |
| Stable Fast 3D | 3D generation | August 2024 | Single-image to 3D mesh in about 0.5 seconds |
| Stable Diffusion 3.5 Large | Image generation | October 2024 | 8B parameter model with improved prompt adherence and quality |
| Stable Diffusion 3.5 Large Turbo | Image generation | October 2024 | Distilled 4-step variant of SD 3.5 Large |
| Stable Diffusion 3.5 Medium | Image generation | October 2024 | 2.5B parameter model optimized for consumer hardware |
| Stable Point Aware 3D (SPAR3D) | 3D generation | January 2025 | Real-time editable 3D mesh reconstruction with point cloud conditioning |
| Stable Virtual Camera | Multi-view video | March 2025 | Transforms 2D images into 3D videos with up to 14 camera paths |
| Stable Audio Open Small | Audio generation | May 2025 | 341M-parameter on-device text-to-audio model for Arm CPUs |
| Stable Audio 2.5 | Audio generation | 2025 | Enterprise-grade audio model with improved quality and control |
| Stable Video 4D 2.0 | 4D generation | 2025 | Upgraded SV4D with improved spatio-temporal consistency |
The Stable Diffusion model family has been the core of Stability AI's product line. Version 1.4 (August 2022) and 1.5 (October 2022) used the original UNet-based latent diffusion architecture. Version 2.0 (November 2022) introduced a new OpenCLIP text encoder and support for higher resolutions, though some users found it less responsive to certain prompting styles than version 1.5.
SDXL 1.0, released in July 2023, represented a major architectural upgrade. It introduced a two-stage pipeline with a base model and an optional refiner, both trained at 1024x1024 resolution. SDXL Turbo, released in November 2023, used a technique called Adversarial Diffusion Distillation (ADD) to compress the generation process to a single step, enabling near-real-time image creation.
In February 2024, Stability AI previewed Stable Cascade, a new text-to-image architecture built on the Würstchen research. The pipeline consists of three smaller models (Stages A, B, and C). Stage C transforms the text prompt into a tiny 24x24 latent image, which Stages A and B then decode into a high-resolution output. By decoupling text-conditional generation from high-resolution decoding, Stability AI reported a 16x training cost reduction compared to comparably sized models. Stable Cascade was released under a non-commercial license.
Stable Diffusion 3 was announced in February 2024 and released as SD 3 Medium on June 12, 2024. It adopted a new Multimodal Diffusion Transformer (MMDiT) architecture, an instance of the broader Diffusion Transformer family, and substantially improved text rendering within images.[32]
The launch was widely regarded as troubled. Community testers immediately noted that SD 3 Medium produced highly distorted human anatomy, particularly in scenes depicting people lying in grass or in other poses requiring full-body composition. Critics attributed the regression to aggressive safety filtering of the training data. The model's original commercial license, which capped commercial usage at 6,000 images per month and required a separate paid agreement for anything beyond that, generated further backlash. Major community hub CivitAI temporarily banned all SD3-derived content over the licensing terms.
Stable Diffusion 3.5, released on October 22, 2024, came in three variants: Large (8 billion parameters), Large Turbo (a 4-step distilled version), and Medium (2.5 billion parameters, released October 29, 2024). The 3.5 family addressed many of the quality complaints raised against SD 3 Medium and was released under Stability AI's new Community License.[10]
The Community License, introduced concurrently with SD 3.5, permits free use (including commercial use) for individuals and organizations with annual revenue below US$1 million, regardless of the source of that revenue. There is no cap on the number of images, videos, or other media that can be generated. Organizations exceeding the $1 million threshold must negotiate a paid Enterprise license. The license was widely viewed as a course correction from the restrictive SD 3 Medium terms.[31]
Stability AI entered the large language model space in April 2023 with StableLM, releasing open-source models at 3 billion and 7 billion parameters. The models were trained on a new experimental dataset built on top of The Pile, an open-source dataset from EleutherAI, with three times more tokens than The Pile alone. In late 2023, the company released Stable LM Zephyr 3B, a 3B-parameter chat model. In January 2024, Stability AI released StableLM 2, a 1.6 billion parameter model trained on 2 trillion tokens of multilingual text data.
Stable Audio 1.0, launched in September 2023, was one of the first commercially viable AI music generation tools, capable of producing 44.1 kHz audio. The audio team, led by Ed Newton-Rex, trained the model on 19,500 hours of licensed music from the AudioSparx content library under a data access agreement. Stable Audio 2.0, released in April 2024, extended generation length to three minutes and introduced audio-to-audio transformation.[13]
Stable Audio Open 1.0 was released on June 5, 2024 as an open-weights model trained on roughly 486,000 audio clips licensed under Creative Commons (CC-0, CC-BY, or CC-Sampling+) from Freesound and the Free Music Archive. It generates stereo 44.1 kHz audio up to 47 seconds in length and is intended for short samples, foley, and instrument loops.[33]
In May 2025, Stability AI and chip designer Arm jointly released Stable Audio Open Small, a 341 million-parameter version of the model distilled and optimized to run entirely on Arm CPUs. The model can generate 11 seconds of audio in under 8 seconds on an Armv9 smartphone, a roughly 30x speedup over standard inference, and is released under the Stability AI Community License.[34]
In 2025, Stability AI released Stable Audio 2.5, its first audio model designed specifically for enterprise use cases, with inference times of under two seconds for full-length three-minute tracks.
Released in November 2023, Stable Video Diffusion (SVD) was Stability AI's first foundation model for video generation. Built on the Stable Diffusion image model, SVD could generate short video clips of 14 or 25 frames at resolutions up to 576x1024 pixels, at frame rates between 3 and 30 frames per second. External evaluations at the time of release found SVD surpassed leading closed-source video models in user preference studies. The model was released as a research preview and was not intended for commercial use at launch.[12]
In July 2024, Stability AI introduced Stable Video 4D (SV4D), the first publicly released video-to-multi-view-video diffusion model. Given an input video, SV4D generates eight novel-view videos following user-specified camera paths, producing five-frame outputs across all eight views in approximately 40 seconds, with full 4D scene optimization taking 20 to 25 minutes. A successor, Stable Video 4D 2.0 (SV4D 2.0), released in 2025, improved spatio-temporal consistency for real-world footage.[35]
Stable Virtual Camera, announced on March 18, 2025, is a multi-view diffusion model that converts one or more 2D images (up to 32) into 3D videos along user-defined camera paths. It supports 14 predefined trajectories, including 360-degree rotation, spirals, and dolly zooms. Released as a research preview under a non-commercial license on Hugging Face, the model is intended for novel-view synthesis without explicit scene reconstruction. A v1.1 checkpoint released in June 2025 fixed foreground-detachment artifacts seen in the initial release.[36]
Stability AI previewed Stable Code Alpha in August 2023, a code-generation language model supporting 18 programming languages including Python, Java, JavaScript, Go, Ruby, and C++. The full Stable Code 3B was released in January 2024 as a 3 billion parameter model with Fill-in-the-Middle (FIM) capabilities. Benchmarks showed it matched or exceeded the code completion quality of models over twice its size, such as Meta's Code Llama 7B.[14]
In March 2024, Stability AI partnered with Tripo AI to release TripoSR, an open-source model for generating 3D meshes from single images. Running on an NVIDIA A100 GPU, TripoSR could produce a draft-quality textured 3D mesh in approximately 0.5 seconds. The model was released under the MIT license and trained on a subset of the Objaverse dataset.[15]
In August 2024, Stability AI released Stable Fast 3D, a successor that generates UV-unwrapped 3D meshes with material parameters and reduced baked-in illumination from a single image in approximately 0.5 seconds.
In January 2025, announced in partnership with NVIDIA at CES, Stability AI introduced Stable Point Aware 3D (SPAR3D). SPAR3D uses a two-stage approach: a lightweight point-diffusion model first generates a sparse 3D point cloud, which is then combined with the input image to produce a detailed mesh. The model supports real-time editing (users can delete, duplicate, stretch, or recolor points in the cloud), produces complete 360-degree geometry including hidden surfaces, generates a final mesh from an edited point cloud in 0.3 seconds, and converts a single image to mesh in 0.7 seconds. SPAR3D is released under the Stability AI Community License, with weights on Hugging Face and a research paper presented at CVPR 2025.[37]
In November 2022, at the AWS re:Invent conference, Stability AI announced it had selected Amazon Web Services (AWS) as its preferred cloud provider. Under this arrangement, Stability AI used Amazon SageMaker for model training, reportedly reducing training time and costs by 58% compared to self-managed infrastructure. The company provisioned one of the largest clusters of machine learning instances on AWS using Amazon Elastic Compute Cloud (EC2). Stability AI also made its open-source models available on Amazon SageMaker JumpStart for AWS customers, and its models are also distributed through Amazon Bedrock.[11]
Stability AI's compute infrastructure included a cluster of more than 4,000 NVIDIA A100 GPUs running on AWS. According to Business Insider, maintaining this system was extremely costly, with the company's operations and cloud costs exceeding $50 million.
Stability AI has maintained a close hardware partnership with NVIDIA. In late 2024, the two companies launched the Stable Diffusion 3.5 NVIDIA NIM microservice, providing optimized inference for enterprise deployment. The SPAR3D launch in January 2025 was announced jointly at CES and tuned for NVIDIA RTX AI PCs.
In March 2025, WPP, the world's largest advertising company, announced a strategic partnership and investment in Stability AI. Through this partnership, WPP gained access to Stability AI's visual media models for image, video, 3D, and audio generation, integrating them into WPP Open, the company's AI-driven operating system. The two companies established a joint R&D pipeline to explore new approaches to brand storytelling, concept testing, and content production. WPP also increased its annual investment in AI and technology to 300 million GBP through WPP Open.[20]
By 2025, Stability AI had formed strategic alliances with both Universal Music Group and Warner Music Group to co-develop fully licensed, commercially safe AI music creation tools for enterprise customers.
On October 23, 2025, Stability AI and video game publisher Electronic Arts announced a strategic partnership to co-develop generative AI models, tools, and workflows for game development. Under the agreement, Stability AI embeds members of its 3D research team directly with EA's artists and developers. Initial focus areas include accelerating the creation of physically based rendering (PBR) materials and 2D textures with color and lighting accuracy preserved across environments, and pre-visualizing complete 3D environments from natural-language prompts. The partnership was Stability AI's most prominent gaming-industry collaboration and followed EA's own $55 billion private equity buyout earlier in 2025.[38]
Despite its high-profile product launches and billion-dollar valuation, Stability AI faced serious financial difficulties during 2023 and 2024. According to Bloomberg, the company was burning approximately $8 million per month as of October 2023, while generating revenue in the low single-digit millions. The company reported losses exceeding $30 million in the first quarter of 2024 alone, with quarterly revenue below $5 million.[22]
The financial strain was compounded by the cost of maintaining its large GPU cluster on AWS. According to various reports, Stability AI owed nearly $100 million to creditors at one point. The company attempted to raise a Series B round at a $4 billion valuation but was unsuccessful.
Revenue grew significantly year over year, rising from approximately $1.5 million in 2022 to $8 million in 2023 and roughly $50 million in 2024. The June 2024 recapitalization eliminated the company's debt and waived $300 million in future obligations, and under new leadership in the second half of 2024 Stability AI focused on API-based revenue and enterprise licensing.[29][30]
Emad Mostaque served as CEO from the company's founding through March 2024. During his tenure, Stability AI grew rapidly and released multiple groundbreaking models. However, his leadership was also marked by controversy.
In June 2023, a Forbes investigation citing more than 30 sources reported that Mostaque had made misleading claims about his educational background, his role at a hedge fund, and the nature of Stability AI's partnership with AWS. The report alleged that Mostaque overstated the AWS relationship as a strategic deal when it was closer to a standard cloud computing contract, and that he had described organizations such as the OECD, WHO, and World Bank as partners, which all three denied. Mostaque contested the findings, saying his Oxford degree was valid but had not been formally conferred because he did not attend his graduation ceremony.[5]
In October 2023, investor Lightspeed Venture Partners sent a letter to Stability AI's board stating that Mostaque's mismanagement had "severely undermined" its confidence in him. Coatue Management separately pushed for Mostaque to resign and launched an internal investigation into his management.[6]
On March 22, 2024, Emad Mostaque resigned as CEO of Stability AI and from his seat on the company's board of directors. In a series of posts on X (formerly Twitter), Mostaque stated that "you can't beat centralized AI with more centralized AI," suggesting he wanted to pursue decentralized AI efforts. However, reporting from Bloomberg, Fortune, and other outlets indicated his departure was driven by sustained pressure from investors and a wave of senior staff departures.[7]
Following Mostaque's resignation on March 23, 2024, Stability AI's board appointed Chief Operating Officer Shan Shan Wong and Chief Technology Officer Christian Laforte as interim co-CEOs while a permanent successor was sought. Wong and Laforte subsequently returned to their previous roles after Akkaraju's June 2024 appointment.[39]
Stability AI experienced a significant exodus of talent during 2023 and 2024. Notable departures included:
| Name | Role | Departure | Notes |
|---|---|---|---|
| David Ha | Chief Researcher | Mid-2023 | Left during broader leadership instability |
| Ed Newton-Rex | VP of Audio | November 2023 | Resigned in protest over the company's fair use stance on copyrighted training data |
| Robin Rombach | Lead Researcher | March 2024 | Original lead developer of Stable Diffusion; co-founded Black Forest Labs |
| Andreas Blattmann | Researcher | March 2024 | Co-author of the Stable Diffusion paper; co-founded Black Forest Labs |
| Dominik Lorenz | Researcher | March 2024 | Co-author of the Stable Diffusion paper |
| Christian Cantrell | VP of Product | 2023/2024 | Left during the broader talent exodus |
| Scott Draves | VP of Engineering | 2023/2024 | Left during the broader talent exodus |
| Patrick Hebron | VP of R&D | 2023/2024 | Left during the broader talent exodus |
| Joe Penna | VP of Applied ML | 2023/2024 | Left during the broader talent exodus |
In April 2024, Stability AI laid off approximately 10% of its global workforce (around 20 employees from a staff of roughly 200), in a move characterized as "right-sizing" the business.[26]
On June 25, 2024, Stability AI appointed Prem Akkaraju as CEO. Akkaraju had previously served as CEO of Weta Digital, the Academy Award-winning visual effects company behind films such as the Avatar series and various Marvel Studios productions. Sean Parker, the entrepreneur and former President of Facebook and Napster co-founder, joined as Executive Chairman.[29]
In September 2024, filmmaker James Cameron joined the Stability AI Board of Directors, bringing high-profile creative industry credibility. Additional board members appointed around this time included Dana Settle, co-founder and managing partner of Greycroft, and Colin Bryant, COO and general partner of Coatue Management.[19]
On August 19, 2024, Stability AI named Hanno Basse as its new Chief Technology Officer. Basse brought more than 30 years of CTO experience from Digital Domain, Microsoft Azure Media and Entertainment, and 20th Century Fox Film Corporation, reinforcing the company's pivot toward film, television, and visual effects use cases.[40]
By December 2024, Akkaraju reported that the company had achieved triple-digit revenue growth, eliminated its debt, and was focused on enterprise customers in film, television, and large-scale content production. The company also achieved SOC 2 Type II and SOC 3 compliance, validating its security controls for enterprise clients.[21]
Stability AI has been involved in several high-profile legal disputes related to the use of copyrighted material in training its AI models.
In February 2023, stock photography company Getty Images filed lawsuits against Stability AI in both the United Kingdom and the United States (in the District of Delaware). Getty alleged that Stability AI had scraped millions of Getty-owned photographs, along with captions and metadata, to train Stable Diffusion without authorization.
On November 4, 2025, Mrs Justice Joanna Smith of the UK High Court handed down judgment in Getty Images (US) Inc & ors v Stability AI Limited [2025] EWHC 2863 (Ch). This was the first UK court decision addressing copyright infringement arising from the training of generative AI models. The court rejected Getty's secondary copyright infringement claims, holding that Stable Diffusion model weights are not "infringing copies" within the meaning of section 27 of the Copyright, Designs and Patents Act 1988, because the model contains statistically learned parameters rather than stored copies or reconstructions of the training images. Importation, possession, or distribution of the model in the UK therefore did not constitute secondary copyright infringement.[16][41]
Getty had earlier abandoned its primary copyright infringement and database right infringement claims after accepting there was no evidence that the training and development of Stable Diffusion took place in the UK. The court did find limited trademark infringement under Sections 10(1) and 10(2) of the UK Trade Marks Act 1994, relating to a small number of generated images that reproduced watermarks similar to Getty's registered marks. The court dismissed Getty's broader Section 10(3) trade mark dilution claim.
The US lawsuit, filed in the District of Delaware in February 2023, was terminated in August 2025.
On January 12, 2023, a group of artists led by Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action copyright infringement lawsuit in the US District Court for the Northern District of California against Stability AI, Midjourney, and DeviantArt. The artists alleged that the defendants used their copyrighted works, scraped from the internet via the LAION dataset, to train AI image generators without permission, and that the AI-generated outputs constituted infringing derivative works.[17]
In October 2023, the court largely granted motions to dismiss by the defendants but allowed direct copyright infringement claims to move forward. After an amended complaint was filed, the court ruled in August 2024 that both direct and induced copyright infringement claims were plausible, allowing the case to proceed to discovery. Plaintiffs filed a third amended complaint on February 27, 2026, with defendants' answers filed on March 13, 2026. The trial is scheduled to begin on September 8, 2026.
In December 2023, a study by David Thiel at the Stanford Internet Observatory found that the LAION-5B dataset, which was used to train early versions of Stable Diffusion, contained 3,226 suspected instances of child sexual abuse material (CSAM), of which 1,008 were externally validated. The LAION-5B dataset had been collected from the public web without consultation with child safety experts and had never been screened against known lists of abusive content.[18]
In response, LAION temporarily took the dataset offline. In August 2024, LAION released Re-LAION-5B, a cleaned version developed in consultation with the Internet Watch Foundation, Human Rights Watch, the Canadian Center for Child Protection, and the Stanford Internet Observatory. In total, 2,236 links to suspected CSAM were removed from the dataset.[28]
Stability AI itself was not the creator of the LAION-5B dataset but had funded and used it for training its early models. The incident intensified scrutiny around data provenance and safety in AI training pipelines.
Stability AI's open-source approach generated significant controversy around content moderation. Unlike other AI image generators that enforced strict filters to prevent the creation of violent, pornographic, or copyright-infringing images, Stability AI's models included only a basic safety filter that could be easily disabled by users running the software locally.
Emad Mostaque publicly advocated for radical openness, arguing that content restrictions would undermine the vision of a democratized AI free from corporate control. Critics countered that this hands-off approach enabled the creation of offensive content, the misuse of the technology against individuals (such as generating non-consensual intimate imagery), and widespread copyright infringement.
In December 2022, Stability AI announced that artists would be able to opt out of the training data for future versions of Stable Diffusion, a concession that the artists' advocacy group Spawning had been pushing for.
The resignation of VP of Audio Ed Newton-Rex in November 2023 brought further attention to these tensions. Newton-Rex stated he could not agree with Stability AI's position that training on copyrighted works constituted "fair use," arguing that generative AI models were clearly being used to create works that compete with the copyrighted material they were trained on.[27]
Stability AI operates in a highly competitive generative AI market. In the image generation space, its primary competitors include:
| Competitor | Model | Type | Notes |
|---|---|---|---|
| OpenAI | DALL-E series | Proprietary | DALL-E 2 (April 2022), DALL-E 3 (October 2023) |
| Midjourney | Midjourney v1 through v7 | Proprietary | Discord-based; known for artistic quality |
| Black Forest Labs | Flux | Open-source / Commercial | Founded by ex-Stability AI researchers Robin Rombach, Andreas Blattmann, and Patrick Esser |
| Google DeepMind | Imagen | Proprietary | Integrated into Google products |
| Adobe | Adobe Firefly | Proprietary | Trained only on licensed/public domain content |
The emergence of Black Forest Labs is particularly notable. Founded in 2024 by the original creators of Stable Diffusion (Robin Rombach, Andreas Blattmann, and Patrick Esser, all of whom left Stability AI), Black Forest Labs released Flux.1 in August 2024. Independent evaluations found Flux comparable to DALL-E 3 in prompt fidelity and to Midjourney 6 in photorealism. By December 2025, Black Forest Labs had raised $300 million in a Series B round at a $3.25 billion valuation, substantially exceeding Stability AI's own peak valuation.[25]
In the audio, video, and language model spaces, Stability AI also competes with companies such as ElevenLabs, Suno, and various open-source LLM providers.
Since Stability AI's core models are open-source and freely downloadable, the company monetizes through several channels:
The Community License, introduced in October 2024, allows free commercial use for businesses with less than $1 million in annual revenue, while larger enterprises must negotiate paid Enterprise terms. There is no cap on the volume of media generated under the Community License.[31]
Under Prem Akkaraju's leadership, Stability AI has focused on stabilizing its finances and building an enterprise-oriented business. Key developments include:
Through early 2026, the company continued to release updated models across image, audio, video, and 3D modalities and to expand its enterprise offering, Stability AI Solutions.
On February 11, 2026, Stability AI announced it had become a full member of the Tech Coalition, the industry alliance that coordinates technology companies against online child sexual exploitation and abuse. The company had spent 2025 in the coalition's Pathways program, a mentorship track for firms building out their child-safety practices, and the full membership marked its move from learning those practices to contributing to them. The step was notable given the LAION-5B CSAM findings that had dogged early versions of Stable Diffusion.[42]
On April 8, 2026, Stability AI launched Brand Studio, a commercial creative-production platform aimed at enterprise marketing and advertising teams rather than developers. The platform centers on Brand Central, where a team trains a custom Brand ID model on its own photography, color palettes, and design elements so that generated output follows brand guidelines. Other components include curated model routing, which automatically picks the best model for a given task and pulls from both Stability's own Stable Diffusion models and third-party models such as Google's Nano Banana and ByteDance's Seedream; a Producer Mode that turns a creative brief into an executable step-by-step production plan; and precision inpainting for product placement and element swaps. Brand Studio is offered in a free Core tier and a paid Enterprise tier, and it fits the company's broader pivot toward selling finished tools to large content producers rather than raw model access.[43]
On May 20, 2026, Stability AI released Stable Audio 3.0, a family of four text-to-audio models trained entirely on licensed data, sourced from the AudioSparx catalog and the openly licensed Freesound library. The lineup consists of a Small SFX model and a Small model (both 459 million parameters, capable of up to two minutes of audio and able to run on-device), a Medium model (1.4 billion parameters), and a Large model (2.7 billion parameters). The Medium and Large models can generate full compositions up to 6 minutes and 20 seconds long while maintaining musical structure, roughly double the length ceiling of Stable Audio 2.0. The Small SFX, Small, and Medium models were released with open weights, while the Large model is available only through the API and paid self-hosting, with companies above $1 million in revenue requiring an Enterprise license.[44][45]