Stability AI

What Is Stability AI?

No company has done more to democratise AI image generation than Stability AI. The 2022 open release of Stable Diffusion fundamentally changed the economics of generative imagery: overnight, anyone with a consumer GPU could generate images at quality levels that previously required proprietary enterprise systems. That decision — to release model weights under a permissive licence — spawned an ecosystem that now encompasses tens of thousands of fine-tuned models, hundreds of community tools, and millions of daily users generating images without ever touching Stability AI's own platform.

Founded in London in 2020 by Emad Mostaque, Stability AI grew rapidly on the back of that model release, raising $181M across multiple rounds. The intervening years were turbulent — delayed payroll in early 2023, leadership restructuring, and Mostaque's departure in March 2024 all tested the company's stability. Prem Akkaraju, former CEO of Weta Digital, took the helm in June 2024 alongside an $80M funding round from Coatue Management and Lightspeed Venture Partners, steadying the ship.

Today, Stability AI operates two parallel tracks. The developer API platform provides managed inference on Stable Diffusion 3.5 and related models, priced by credit. The open-weight releases continue under permissive licences, fuelling an ecosystem that Stability AI does not directly monetise but benefits from enormously in brand and distribution. With approximately 191 employees and $104M in annual revenue, the company is lean relative to its market influence.

Key Features

Stable Diffusion 3.5 Models

SD 3.5 is Stability AI's current flagship model family, built on the Multimodal Diffusion Transformer (MMDiT) architecture that processes image and language tokens separately — improving text rendering and prompt adherence significantly over earlier SD versions. The family spans three tiers:

SD 3.5 Large uses 8 billion parameters and delivers the highest output quality. Images are generated at up to 1 megapixel resolution with exceptional prompt fidelity, complex composition handling, and reliable text-in-image rendering. At 6.5 credits per image (~$0.065), it is the premium option for quality-sensitive production work.

SD 3.5 Medium is the efficiency optimised model, balancing speed and quality for high-volume pipelines. At 3.5 credits (~$0.035/image), it undercuts most cloud competitors on per-image cost.

SD 3.5 Large Turbo achieves near-Large quality in just 4 inference steps — roughly 4 seconds on platform infrastructure — making it practical for near-real-time applications at 4 credits per image.

Inpainting, Outpainting, and Depth Control

Beyond text-to-image, the API exposes image editing primitives that are essential for production workflows. Inpainting lets you modify specific masked areas of an existing image while preserving surrounding context. Outpainting expands images beyond their original canvas — useful for creating banner assets or repurposing portrait images for landscape formats. Depth-to-image uses structural guidance from a reference image to control the spatial layout of generated output.

Stable Image Ultra

The API's premium offering — 8 credits per image (~$0.08) — Stable Image Ultra represents Stability AI's highest quality tier, targeting commercial outputs where detail, coherence, and print resolution matter. This sits above SD 3.5 Large in the quality hierarchy and is positioned against Midjourney's top tier and DALL-E 3.

Open-Weight Ecosystem

The most strategically important feature has no API endpoint. Stability AI's open-weight releases — available on Hugging Face under Apache 2.0 and similar licences — enable organisations to deploy models entirely on their own infrastructure. This means zero API costs at scale, full control over data flows, and the ability to fine-tune on proprietary imagery datasets. The ecosystem built on these releases (ComfyUI, AUTOMATIC1111, thousands of community LoRAs) vastly amplifies the practical utility of Stability AI's core models.

Pricing

Stability AI's credit system is straightforward: $10 buys 1,000 credits, and credits are consumed per image based on the model selected. For teams doing occasional generation, the pay-as-you-go model with 25 free trial credits on sign-up requires no upfront commitment.

For production workflows, the math favours Stability AI. At $0.035/image for SD 3.5 Medium, generating 10,000 images costs $350. Midjourney's basic plan at $10/month caps at roughly 200 images, making the cost-per-image approximately $0.05. DALL-E 3 via the OpenAI API runs at $0.04–$0.12/image depending on resolution and quality settings.

The open-weight alternative changes the calculus entirely: organisations running SD 3.5 Large on their own GPU infrastructure pay only compute costs, which can reduce per-image cost to fractions of a cent at scale. For EU enterprises concerned about data sovereignty, self-hosting delivers both cost efficiency and compliance simultaneously.

Enterprise pricing is custom and includes volume discounts, SLA guarantees, dedicated support, and on-premise deployment options. Given Stability AI's 191-person team servicing a massive developer base, the value of enterprise support contracts depends heavily on your negotiating position.

EU Compliance & Privacy

This is where an honest assessment requires nuance. Stability AI is a UK company (post-Brexit), not an EU company. The hosted API platform does not carry EU data sovereignty guarantees — the UK's adequacy decision with the EU covers personal data transfers but does not make the UK equivalent to an EU data processor for all purposes. Organisations subject to strict EU data localisation requirements should not assume the hosted API qualifies.

The platform has achieved SOC 2 Type II and SOC 3 certification — meaningful security assurance for enterprise procurement. These certifications validate controls around availability, security, and confidentiality, providing documented audit evidence that compliance teams can review.

The clearest path to EU compliance is the open-weight self-hosting route. Running SD 3.5 models on EU-hosted GPU infrastructure (Hetzner, OVHcloud, Deutsche Telekom) gives organisations complete control: no external data transfer, no third-party API dependency, and full GDPR alignment. For sensitive use cases — medical imaging, biometric content, regulated industries — this is the only appropriate deployment model regardless of Stability AI's certifications.

Who It's Best For

EU software teams building AI-powered creative tools who need high-quality image generation at scale with cost efficiency and the option to self-host.

Creative agencies doing high-volume content production who need per-image cost control and the flexibility to use different model qualities for different output tiers.

AI engineers and researchers who want to fine-tune models on proprietary datasets — the open-weight releases and active ecosystem make Stability AI the best foundation for custom model development.

Regulated organisations that need generative AI but cannot send data to US or UK-hosted APIs — self-hosted SD 3.5 on EU infrastructure is the path that satisfies both requirements.

The Verdict

Stability AI occupies a genuinely unique position in the AI image generation market. The combination of a capable hosted API and open-weight model releases provides flexibility that closed competitors simply cannot match. SD 3.5 is technically excellent, and the credit-based pricing competes well against Midjourney and DALL-E at equivalent quality levels. The caveats matter: UK headquarters means no EU data sovereignty on the hosted API, the company's history of financial instability warrants monitoring, and enterprise support is thin relative to the developer community's size. For EU teams requiring data sovereignty, self-hosting open-weight models is both the compliance and the cost-efficiency answer.

Frequently Asked Questions

Are Stable Diffusion models free to use commercially?

Yes. Stability AI releases Stable Diffusion models under open licences (typically Apache 2.0) that permit commercial use without royalties. The API platform also permits commercial use under its terms of service. Check the specific licence file for each model version, as terms occasionally vary between releases.

Is Stability AI suitable for EU-regulated industries?

The hosted API is a UK-based service without EU data sovereignty guarantees, making it unsuitable for regulated EU industries that require data localisation. The open-weight model releases are the better path — deploying SD 3.5 on EU-based GPU infrastructure (Hetzner, OVHcloud, or your own colocation) keeps all data within the EU under your complete control.

How does Stable Diffusion 3.5 compare to Midjourney?

SD 3.5 Large is competitive with Midjourney v6 on photorealistic and compositional quality, but Midjourney generally produces more aesthetically polished outputs by default, with less prompt engineering required. SD 3.5's advantages are API access, programmatic control, lower cost per image at volume, and the ability to self-host. Midjourney has no self-hosting option and requires Discord or its web app.

What tools does the Stable Diffusion ecosystem include?

The community around Stability AI's open-weight releases has produced ComfyUI (node-based workflow editor), AUTOMATIC1111 (browser-based WebUI), thousands of community LoRA fine-tunes on Civitai, and integrations into platforms like Adobe Photoshop, Canva, and Leonardo.ai. This ecosystem multiplies the practical utility of Stability AI's models far beyond what the company itself provides.

What happened to Stability AI's leadership and finances?

Founder Emad Mostaque resigned in March 2024 following reports of financial difficulties including delayed payroll. Prem Akkaraju (former Weta Digital CEO) was appointed CEO in June 2024, accompanied by an $80M funding round from Coatue and Lightspeed. The company has since stabilised, now reporting approximately $104M in annual revenue with around 191 employees.