Stable Diffusion is a family of open-weight text-to-image models from Stability AI, and teams that use it have to make choices about versions, licences, hosting and provenance that closed image services make for them. The first public Stable Diffusion release was announced by Stability AI on 22 August 2022; the current flagship generation, Stable Diffusion 3.5, arrived on 22 October 2024 (Large and Large Turbo) and 29 October 2024 (Medium). This page sets out the model line-up, the Stability AI Community License and its USD 1M revenue threshold, API credit prices, self-hosting requirements, provenance and data handling, and the dated release and deprecation history. It is written for designers and content teams, developers, platform engineers and the security, legal and privacy teams who approve generative AI tools.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- Current generation is SD 3.5. Stable Diffusion 3.5 Large (8.1 billion parameters), Large Turbo (a 4-step distilled Large) and Medium (2.5 billion parameters) are the current open-weight models.
- Free below USD 1M revenue. The Stability AI Community License covers research, non-commercial and commercial use for individuals and organisations with annual revenue under USD 1M; above that an Enterprise License is required.
- SDXL uses a different licence. SDXL 1.0 is released under the CreativeML Open RAIL++-M License, not the Community License, so check the licence per model.
- API is priced in credits. One credit costs USD 0.01; SD 3.5 Large costs 6.5 credits per generation and SD 3.5 Medium 3.5 credits, per Stability AI's pricing page.
- Self-hosted outputs carry no automatic provenance. Stability AI's release notes list C2PA signing on its hosted API services; images you generate on your own hardware have only what your pipeline adds.
What is Stable Diffusion?
Stable Diffusion is a latent diffusion model family: it generates images by removing noise step by step in a compressed latent space, guided by a text prompt that one or more text encoders turn into embeddings. The earlier generations (Stable Diffusion 1.x, 2.x and SDXL) use a U-Net denoiser. Stable Diffusion 3 and 3.5 replace it with a Multimodal Diffusion Transformer (MMDiT), which Stability AI's model card for SD 3.5 Large says improves image quality, typography, complex prompt understanding and resource efficiency. What sets Stable Diffusion apart from hosted-only image services is that Stability AI publishes the model weights on Hugging Face, so anyone who accepts the licence can download, run, fine-tune or build LoRAs on the models on their own hardware.
The family is also the basis of Stability AI's hosted products: the Stability Platform API, the Stable Image services (Ultra and Core), and the web-based Stable Assistant. Stable Image Ultra moved to SD 3.5 Large on 29 October 2024, according to the API release notes.
Key facts
| Field | Value (as of 2026-11-15) |
|---|---|
| Provider | Stability AI |
| Modality | Text-to-image and image-to-image; editing, upscaling and control services on the API |
| Current flagship | Stable Diffusion 3.5 Large, 8.1 billion parameters, MMDiT |
| Text encoders (SD 3.5) | OpenCLIP-ViT/G, CLIP-ViT/L and T5-xxl |
| Open weights | Yes, on Hugging Face, with inference code on GitHub |
| Licence (SD 3.5) | Stability AI Community License; Enterprise License above USD 1M annual revenue |
| Licence (SDXL 1.0) | CreativeML Open RAIL++-M License |
| Output ownership | Users own outputs, subject to law and the Acceptable Use Policy |
| Hosted access | Stability Platform API, Stable Assistant, Amazon Bedrock and other partners |
| Provenance | C2PA signing listed in API release notes for hosted services |
Models and versions
The table lists the models Stability AI currently presents on its Stable Image page and model cards. Parameter counts are as Stability AI states them; the SD 3.5 announcement gives 8.1 billion for Large, while the Hugging Face card rounds it to 8 billion.
| Model | Parameters | Architecture | Positioning (Stability AI) | Licence |
|---|---|---|---|---|
| Stable Diffusion 3.5 Large | 8.1B | MMDiT | Professional use at 1 megapixel | Community License |
| Stable Diffusion 3.5 Large Turbo | Distilled from Large | MMDiT | High-quality images in 4 steps | Community License |
| Stable Diffusion 3.5 Medium | 2.5B | MMDiT-X | Consumer hardware, 0.25 to 2 megapixels | Community License |
| Stable Diffusion 3 Medium | 2B | MMDiT | First SD3 open release (June 2024) | Released under a Stability community licence |
| Stable Diffusion XL 1.0 | 3.5B base, 6.6B with refiner | U-Net latent diffusion, base plus refiner | Native 1024x1024 output | CreativeML Open RAIL++-M |
| Stable Diffusion XL Turbo | Distilled from SDXL | U-Net | Images in as few as one step | Community License |
As of 2026-11-15. SD 3.5 Flash is listed on the API pricing page; Stability AI's model pages do not describe it further.
SDXL ensemble: the SDXL base model can run alone or hand its noisy latents to a separate refiner model for the final denoising steps. The model card warns that SDXL cannot render legible text and struggles with compositional prompts, two areas the MMDiT generation targets.
SD 3.5 Medium resolutions: the Medium model card describes progressive training from 256 up to 1440 pixels, with a final stage of mixed-scale training for multi-resolution generation, and warns of edge artefacts when the T5 prompt exceeds 256 tokens.
Pricing and plans
The weights are free to download under the Community License for eligible users; costs are hardware and electricity. The hosted API is billed in credits, and Stability AI states that one credit equals USD 0.01 and that pricing is subject to change.
| API service | Credits per call | USD equivalent |
|---|---|---|
| Stable Image Ultra | 8 | 0.08 |
| SD 3.5 Large | 6.5 | 0.065 |
| SD 3.5 Large Turbo | 4 | 0.04 |
| SD 3.5 Medium | 3.5 | 0.035 |
| SD 3.5 Flash | 2.5 | 0.025 |
| Stable Image Core | 3 | 0.03 |
| SDXL 1.0 | From 0.9 | From 0.009 |
| Creative / Conservative / Fast Upscaler | 60 / 40 / 2 | 0.60 / 0.40 / 0.02 |
| Inpaint / Outpaint / Erase Object | 5 / 4 / 5 | 0.05 / 0.04 / 0.05 |
As of 2026-11-15, from docs.stability.ai/pricing. USD figures are credits multiplied by USD 0.01. Enterprise licence pricing is not published; Stability AI quotes it on request. Pricing on cloud marketplaces is set by each cloud.
How does the Stability AI licence work?
Licensing is the question most teams need answered first, because it differs by model and by the size of the organisation.
| Licence | Who it covers | Key terms | Models |
|---|---|---|---|
| Stability AI Community License | Researchers, developers, small businesses and creators with annual revenue under USD 1M, from any source | Free research, non-commercial and commercial use; derivatives such as fine-tunes and LoRAs may be distributed; revocable on breach | Core Models, including the SD 3.5 suite and SDXL Turbo |
| Enterprise License | Organisations and API providers above USD 1M annual revenue | Custom pricing; commercial use and implementation support; optional custom training | Core Models |
| CreativeML Open RAIL++-M | Any user who accepts its use restrictions | Responsible AI licence with use-based restrictions | SDXL 1.0 base and refiner |
| CreativeML OpenRAIL-M | Any user who accepts its use restrictions | Described at launch as permissive for commercial and non-commercial use | Original 2022 Stable Diffusion release |
As of 2026-11-15, from stability.ai/license and the model cards.
- Threshold trigger: organisations that reach USD 1M annual revenue must contact Stability AI to buy an Enterprise License. Commercial use of research work requires registration.
- Derived foundation models: creating a new foundational model architecture from a Core Model is restricted; LoRAs, hypernetworks, fine-tunes and retrains are not treated as foundation models.
- Acceptable Use Policy: applies to every licence. It prohibits, among other things, child sexual abuse material, non-consensual intimate imagery, impersonation, misleading users about the nature of outputs and failing to disclose AI interaction where it is not apparent.
See the Open-Weight vs Closed Models Explained for what "open weights" does and does not mean.
Features
- Generate: text-to-image across the SD 3.5 models and Stable Image Ultra and Core; Ultra gained image-to-image on 7 November 2024.
- Edit: erase object, inpaint, outpaint, remove background, search and recolor, search and replace, and replace background and relight.
- Upscale: Creative, Conservative and Fast upscalers.
- Control: sketch, structure, style guide and style transfer, which condition generation on an input image.
- Customise: with open weights, fine-tuning and LoRA training on your own data; Stability AI publishes a fine-tuning tutorial for SD 3.5 Large.
Availability
| Route | What you get | Who controls data |
|---|---|---|
| Self-hosted weights (Hugging Face, GitHub) | Full model, local inference via Diffusers, ComfyUI or the reference code | You |
| Stability Platform API | SD 3.5, Stable Image services, edit, upscale and control endpoints | Stability AI, under its terms and privacy policy |
| Amazon Bedrock | Stable Image Ultra, SD 3.5 Large, Stable Image Core and Stability AI Image Services | AWS and your account configuration |
| Other hosts named by Stability AI | Replicate, Fireworks AI, DeepInfra and ComfyUI for SD 3.5 | The host, under its terms |
| Stable Assistant | Web app for non-developers | Stability AI |
As of 2026-11-15. AWS notes that support for other Stability AI models on Bedrock is being deprecated.
Hardware: Stability AI states SD 3.5 Medium needs 9.9 GB of VRAM excluding text encoders and runs out of the box on consumer hardware; the original 2022 model needed 6.9 GB at release. No VRAM figure is published for SD 3.5 Large.
A practical self-hosting sequence:
- Confirm licence eligibility (revenue under USD 1M or an Enterprise License) and record the decision.
- Accept the licence on Hugging Face with an organisation account, not a personal one, and download weights in safetensors format.
- Pin model and library versions (for example the Diffusers release named in the model card) and verify file hashes.
- Run inference in an isolated environment with no inbound internet exposure; put an authenticated gateway in front of any shared service.
- Add prompt and output filtering, because Stability AI's model card advises developers to apply their own mitigations and guardrails.
- Add provenance metadata or labelling to outputs in your pipeline, and log prompts and outputs according to your retention policy.
See the Open-Weight Model Security Guide and the Hugging Face Security Guide for supply chain and token controls.
Hosted or self-hosted: which route fits?
The choice is a set of factual trade-offs rather than a ranking. The criteria below come from the licence, pricing and privacy terms set out above.
| Criterion | Self-hosted weights | Hosted API or cloud |
|---|---|---|
| Data location and training use | Prompts and images stay on your infrastructure | Governed by the provider's privacy policy and any enterprise contract |
| Cost model | GPU capacity, power and operations staff | Per-call credits or cloud per-image pricing |
| Provenance | Only what your pipeline adds | C2PA signing listed on Stability AI hosted services |
| Safety filtering | Your responsibility, per the model card | Provider filters prompts and outputs |
| Model choice | Any released weights, plus fine-tunes and LoRAs | Models the provider currently serves; older ones are deprecated on notice |
| Licence | Community or Enterprise License applies directly to you | Provider terms of service apply to API use |
Small teams: a hosted API avoids GPU operations and gives provenance signing by default, at the cost of sending prompts to a third party. Larger organisations with sensitive inputs, high volumes or a need for custom fine-tunes often weigh self-hosting, but must then supply filtering, labelling, access control and patching themselves, and confirm Enterprise licensing once revenue passes USD 1M.
Provenance and labelling
Stability AI's safety page says it is implementing C2PA content authenticity standards so users and platforms can identify AI-assisted content made with its hosted services, and that it works with the Content Authenticity Initiative. Its API release notes list C2PA signing as a feature on several hosted services from April 2024 onwards. That applies to hosted generation. Weights you run yourself add no metadata unless your pipeline does, and C2PA metadata is removed by many re-encoding and upload paths. Stability AI's safety page does not describe an invisible watermark on outputs. See the AI Content Provenance and C2PA Guide for how to sign and verify content.
Data and privacy
- Training use: Stability AI's privacy policy says inputs and outputs may be used to train, fine-tune and evaluate models, with an opt-out in account settings.
- Retention: no fixed period; data is kept while the account is active and needed to provide the service, subject to agreement and law.
- Location: data may be processed outside the EEA, Switzerland and the UK, including in the US.
- Shared workspaces: inputs and outputs are visible to members of a shared enterprise account or workspace.
- Self-hosting: prompts and images stay on your infrastructure; your own retention and access policies apply.
- API keys: treat Stability Platform API keys as secrets, store them in a vault and rotate them. See the Self-Hosted LLM Security Guide for comparable controls on self-hosted models.
Release history
| Date | Event |
|---|---|
| 22 August 2022 | Public release of Stable Diffusion under CreativeML OpenRAIL-M |
| 26 July 2023 | SDXL 1.0 announced, CreativeML Open RAIL++-M |
| 21 March 2024 | Stable Image services (Generate, Upscale, Edit, Control) on the API |
| 5 June 2024 | Stable Image Ultra on the API |
| 12 June 2024 | Stable Diffusion 3 Medium released |
| 11 October 2024 | Legacy API models deprecated, including stable-diffusion-512-v2-1 and stable-diffusion-xl-1024-v0-9 |
| 22 October 2024 | Stable Diffusion 3.5 Large and Large Turbo |
| 29 October 2024 | Stable Diffusion 3.5 Medium; Stable Image Ultra moves to SD 3.5 Large |
| 17 April 2025 | SD 3.0 APIs deprecated; requests route to SD 3.5 equivalents |
| 24 July 2025 | Stable Video Diffusion API and Stable Diffusion 1.6 API no longer supported |
As of 2026-11-15, from Stability AI announcements and docs.stability.ai release notes.
Limits
- Not factual: model cards state the models were not trained to be factual or true representations of people or events.
- Safety not guaranteed: training data was filtered, but Stability AI says this does not guarantee all harmful content was removed; testing was mainly in English.
- Bias: the SDXL card notes that image generation can reinforce or exacerbate social biases.
- Licence revocation: the Community License can be revoked if its terms are breached.
- API churn: older endpoints are deprecated on dated notices, so pin model names and watch the release notes.
Practitioner checklist
- Record which Stable Diffusion model and licence each team uses, and confirm revenue eligibility for the Community License.
- Download weights only from Stability AI's official Hugging Face organisation and verify hashes.
- Decide hosted or self-hosted per use case, based on data sensitivity and the privacy policy's training clause.
- Opt out of training use in account settings where hosted data must not train models.
- Store API keys in a secrets manager, scope them per application and rotate them.
- Apply prompt and output filters to self-hosted deployments, as the model card advises.
- Preserve C2PA metadata from hosted services and add labelling for self-hosted outputs where disclosure rules apply.
- Monitor Stability AI release notes for deprecations and update pinned model names.
Standards and references
- Stability AI: License
- Stability AI: Introducing Stable Diffusion 3.5
- Stability AI: Stable Diffusion 3.5 Large model card
- Stability AI: Stable Diffusion 3.5 Medium model card
- Stability AI: SDXL 1.0 base model card
- Stability AI: API pricing
- Stability AI: API release notes
- Stability AI: Safety
- Stability AI: Privacy Policy
- AWS: Stability AI models in Amazon Bedrock
Related NHI Mgmt Group resources: Open-Weight Model Security Guide · AI Content Provenance and C2PA Guide · AI Image and Video Generation at Work · Hugging Face Security Guide