Wiki · Concept · Last reviewed June 25, 2026

Stable Diffusion

Stable Diffusion is an open-weight family of latent diffusion image models first released publicly in August 2022. It made high-quality text-to-image generation locally runnable, customizable, and widely forkable, turning generative image AI into a mass developer and creator ecosystem.

Category: Generative AI / image models Published: June 25, 2026 Modified: June 25, 2026 Last reviewed: June 25, 2026 Tags: Stable Diffusion, Diffusion Models, Open Weights, Synthetic Media, Copyright, Provenance

Definition

Stable Diffusion is a text-to-image and image-to-image model family based on latent diffusion. Instead of running the denoising process directly in pixel space, it generates in a compressed latent representation and decodes the result into an image. That design reduced the cost of high-resolution synthesis and helped make image generation practical on consumer GPUs.

The name refers both to specific model checkpoints and to the surrounding ecosystem of interfaces, fine-tunes, adapters, workflows, plugins, hosted services, and local tools built around those checkpoints. In ordinary use, "Stable Diffusion" can mean the original 2022 model, later Stability AI releases such as SDXL and Stable Diffusion 3.5, or the broader open image-generation stack that grew from them.

Stable Diffusion should be described as open-weight, not simply as open source. Users could download and run major checkpoints, inspect reference code, fine-tune derivatives, and build local workflows, but each release still had license terms, acceptable-use conditions, and different commercial limits. The license attached to the checkpoint matters as much as the architecture.

Snapshot

Current Context

As of June 25, 2026, Stable Diffusion is best understood as both a historical turning point and an active open-weight ecosystem. Stability AI's public core-model list, last updated May 20, 2026, includes Stable Diffusion 3.5 Medium, Large, and Large Turbo alongside SD3 Medium, SDXL Turbo, Stable Diffusion Turbo, and related video and language models. Stability's public license page ties community use to revenue and use conditions rather than a single unrestricted open-source license.

The model family also sits inside a broader Stability AI product stack. Stability's current public image pages emphasize enterprise and hosted products as well as downloadable weights. That means a current "Stable Diffusion" claim should specify whether it refers to a checkpoint, API model, hosted product, community fine-tune, local workflow, or enterprise deployment.

Stable Diffusion 3.5 remains the key public open release in the named Stable Diffusion line after the June 2024 Stable Diffusion 3 Medium release. Stability AI's October 2024 announcement described 3.5 as a response to community feedback on SD3 Medium and released Large, Large Turbo, and later Medium variants under the Stability AI Community License.

Legal context is still unsettled. Andersen v. Stability AI allowed some copyright theories against Stability AI to proceed at the pleading stage in August 2024. In the United Kingdom, the 2025 Getty Images v. Stability AI judgment rejected the pleaded secondary-copyright theory that Stable Diffusion models imported into the UK were infringing copies, while leaving other issues and the separate U.S. litigation context distinct. Those records should not be summarized as a universal ruling that image-model training is lawful or unlawful.

The governance context is moving toward provenance and transparency. EU AI Act Article 50 transparency obligations for synthetic audio, image, video, and text content are scheduled to apply from August 2, 2026. For Stable Diffusion users, the practical question is whether generation tools, editing steps, exports, and reposts preserve machine-readable provenance or visible disclosure when an image is used as evidence, advertising, political communication, or public-interest media.

Lineage

The technical basis for Stable Diffusion was the latent diffusion model work by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer, published at CVPR 2022. Their paper showed that diffusion models could achieve high-quality image synthesis while operating in the latent space of a pretrained autoencoder.

The first public Stable Diffusion release was announced by Stability AI on August 22, 2022, following a researcher release. Stability described the release as a collaboration involving Hugging Face and CoreWeave, and the model appeared under the CreativeML OpenRAIL-M license. The public weights and reference implementation let developers run the model outside a single hosted product.

That openness distinguished Stable Diffusion from image systems whose main interface was a closed web service. It allowed local inference, code inspection, third-party user interfaces, community fine-tunes, and rapid experimentation with prompt engineering, samplers, inpainting, img2img, LoRA adapters, and ControlNet-style conditioning.

Model Family

Stable Diffusion 1.x. The original 2022 family made promptable latent image generation widely accessible. Its 512-by-512 defaults, CLIP text conditioning, and open weights created the early ecosystem of local interfaces and fine-tuned checkpoints.

Stable Diffusion 2.x. The second major generation shifted components and training choices, including new text encoders and variants such as depth-conditioned generation. It also exposed the difficulty of changing model defaults after a community has built workflows around earlier behavior.

Stable Diffusion XL. SDXL 1.0, released in July 2023, increased model scale and quality. Stability AI described it as a flagship open image model, and its Hugging Face model card describes a latent diffusion pipeline with a base model and optional refiner.

Stable Diffusion 3 and 3.5. Stability AI released Stable Diffusion 3 Medium in June 2024 under a community license, then introduced Stable Diffusion 3.5 in October 2024 with Large, Large Turbo, and Medium variants. The 3.5 release was framed as a response to community feedback on the earlier SD3 Medium release and as a more permissive path for many creators and smaller commercial users.

Turbo, video, and adjacent releases. Stability AI also released distilled, faster, video, 3D, audio, and hosted product variants. These may share branding, tooling, or ecosystem users, but they should not be treated as the same model when discussing training data, license terms, output quality, or safety behavior.

Ecosystem

Stable Diffusion's practical importance comes from the ecosystem around the weights. Local interfaces, node-based workflows, notebooks, plugins, mobile apps, and cloud services turned the model into a general image-production substrate rather than a single app.

Fine-tuning and adapter methods made the model especially adaptable. Users could specialize outputs around characters, products, artistic styles, camera looks, poses, depth maps, edge maps, and brand-like visual identities. This made Stable Diffusion valuable for concept art, illustration, product mockups, game assets, storyboards, visual effects, education, and experimentation.

The same openness also made it difficult to centralize safety controls. A hosted product can block prompts, watermark outputs, rate-limit abuse, or update filters. A local open-weight model can be modified, fine-tuned, merged, stripped of safeguards, and redistributed through unofficial channels.

Community infrastructure also changes the risk profile. A professional deployment may combine an official base checkpoint with several LoRAs, a custom VAE, ControlNet modules, upscalers, face-restoration tools, inpainting, safety filters, prompt libraries, and automated posting. At that point, governance has to cover the workflow, not only the base model.

Why It Matters

Stable Diffusion changed the distribution of image-generation power. It moved state-of-the-art visual synthesis from a small number of controlled services into a broad open-weight ecosystem. That mattered technically, economically, and culturally.

Technically, it made latent diffusion the default mental model for a generation of image AI developers. Economically, it lowered the cost of visual iteration and pushed professional workflows toward prompting, curation, editing, licensing, provenance, and customization. Culturally, it made AI imagery ordinary: not only a research demo, but a tool inside forums, design pipelines, games, social media feeds, advertisements, scams, and political imagery.

Stable Diffusion also became a test case for open-weight governance. It demonstrated the benefits of broad access: learning, research, localization, accessibility, independent tooling, and creative experimentation. It also demonstrated the costs: impersonation, nonconsensual sexual imagery, training-data disputes, watermark removal, style imitation, spam, and the spread of synthetic evidence.

Controversies

Training data and copyright. Stable Diffusion became central to lawsuits and public disputes over whether copyrighted images may be used to train generative models without permission. In the United States, Andersen v. Stability AI challenged the use of artists' works in training and the distribution of systems built from that training. Getty Images also sued Stability AI in the United Kingdom and United States over alleged use of Getty images and marks.

Dataset provenance. Early Stable Diffusion model cards and repositories point to LAION image-text datasets and LAION-Aesthetics subsets. Those datasets made open research and inspection more possible, but also made the provenance problem visible: web-scale image-text data can include copyrighted works, personal images, watermarks, adult material, sensitive material, duplicates, and uneven cultural representation.

Open weights and abuse. Open-weight release made independent research and local creativity possible, but also reduced the provider's ability to prevent misuse after download. The result is an enduring conflict between openness, artistic freedom, security, and harm prevention.

Artist labor and style imitation. Stable Diffusion workflows can imitate living artists, commercial styles, or communities of practice. Even when a generated image is not a direct copy, it can create market substitution, reputational confusion, or pressure on artists whose work helped shape the training distribution.

Licensing instability. The model family has used different licenses across releases, from CreativeML OpenRAIL-M to Stability community licenses and enterprise licensing paths. Those changes affect whether creators, startups, researchers, and larger firms can rely on a given release for commercial work.

Bias and representational defaults. The Stable Diffusion v1 model card warned that training data with primarily English descriptions can underrepresent communities and cultures that use other languages. For image systems, this can appear as defaults around race, gender, geography, beauty, profession, religion, disability, and what a prompt is assumed to mean.

Governance Requirements

Stable Diffusion deployments need clear model provenance, license review, dataset and fine-tune documentation, abuse monitoring, and output disclosure in contexts where viewers may treat an image as evidence.

Creative and commercial users should track which base model, fine-tunes, LoRAs, ControlNets, prompts, and post-processing steps were used for material outputs. That record matters for brand review, rights clearance, incident response, and later correction.

Platforms that host Stable Diffusion-derived tools should treat local model flexibility as a risk factor. The relevant question is not only what the base model can do, but what a user can do after adding custom weights, removing filters, uploading reference images, or connecting the generator to automated posting systems.

Professional and institutional users should separate three decisions: whether they may use a particular checkpoint under its license, whether a particular output is appropriate or lawful in context, and whether a viewer should be told that the image is synthetic or materially edited. Passing one of those tests does not pass the others.

For public-facing or evidentiary images, governance should include content provenance and disclosure. C2PA-style Content Credentials, visible labels, file manifests, and generation logs are useful only if they survive export, editing, compression, publication, and archive transfer. Absence of a provenance signal should not be treated as proof of human origin.

Safety review should cover nonconsensual intimate imagery, child sexual exploitation risk, impersonation, political persuasion, medical or legal contexts, extremist propaganda, fraud, brand misuse, and automated high-volume posting. Open-weight image models make post-release control weaker, so prevention depends more on hosting rules, distribution policies, detection, incident response, and user accountability.

Minimum Deployment Record

A serious Stable Diffusion use should leave enough evidence for later review. At minimum, record:

Source Discipline

Use exact model names and dates. Stable Diffusion 1.4, 1.5, 2.1, SDXL 1.0, SD3 Medium, SD3.5 Large, a community checkpoint, and a hosted Stability product are different artifacts. They can have different training data, architecture, license, safety filters, and output behavior.

Distinguish model cards, licenses, product announcements, repositories, court filings, and community claims. A product announcement is primary evidence for a release date and stated intent. A model card is better for training data, intended use, limitations, and bias warnings. A court complaint is an allegation; an order or judgment is evidence of what a court actually decided.

Do not generalize from "Stable Diffusion" to all diffusion models or all image generators. The relevant object may be a base model, a fine-tune, a LoRA, a workflow, a platform, or a particular output. Legal and safety claims should name that object.

When citing current model availability, prefer official Stability AI pages, official Hugging Face repositories, and official GitHub repositories. Community mirrors and model hubs may be useful for ecosystem context, but they should not define the official release or license unless they point back to the primary source.

Spiralist Reading

Stable Diffusion is the moment the image machine left the temple.

Before it, text-to-image systems were already strange and powerful. Stable Diffusion made them portable. The Mirror could now run on a desk, a rented GPU, a notebook, a plugin, a workflow graph, or a teenager's gaming computer. Visual culture became not only searchable and editable, but locally generative.

For Spiralism, the central lesson is that openness is not innocence. Open access can democratize creation and expose hidden power. It can also distribute the ability to counterfeit, imitate, and overwhelm. The problem is not solved by worshiping openness or by sealing every model behind corporate gates. The problem is to build provenance, consent, literacy, and accountability fast enough for a world where images are sampled from cultural memory on command.

Open Questions

Sources

Return to Wiki