Stable Diffusion

What is Stable Diffusion?

Stable Diffusion is Stability AI’s family of image generation models that create pictures from text prompts and existing images. The current line is Stable Diffusion 3.5: Large has 8.1 billion parameters and generates at 1 megapixel, Large Turbo produces images in four steps, and Medium has 2.5 billion parameters for consumer hardware. Older SDXL and SDXL Turbo models remain available as legacy options, and directories such as There’s An AI For That list 16 Stable Diffusion releases in total.

The weights are published on Hugging Face with inference code on GitHub, so anyone can run the models on their own GPU. Use is governed by the Stability AI Community License, which is free for research, non-commercial use and businesses under $1 million in annual revenue; larger firms need an Enterprise Licence. You can also reach the models through the Stability Platform API, cloud partners such as Amazon Bedrock and Microsoft Azure AI Foundry, or the Stable Assistant app. The old DreamStudio address now redirects to Brand Studio, Stability AI’s newer commercial creative platform.

Key Features

• Stable Diffusion 3.5 Large: Stable Diffusion 3.5 Large is the family’s flagship, with 8.1 billion parameters and 1 megapixel output. Stability AI describes it as its most powerful image model for quality and prompt adherence, aimed at professional work.
• Large Turbo for speed: Stable Diffusion 3.5 Large Turbo is a distilled version that generates images in just four steps. It trades a little detail for much faster output, which suits rapid concept iteration and high-volume API jobs.
• Medium for consumer GPUs: Stable Diffusion 3.5 Medium uses 2.5 billion parameters and needs about 9.9 GB of VRAM, excluding text encoders. It generates between 0.25 and 2 megapixels, so hobbyists can run it on a single gaming graphics card.
• Open weights: Stable Diffusion weights download from Hugging Face, with inference code on GitHub. You can fine-tune the models, build derivative versions and run everything offline, which closed services such as Midjourney do not allow.
• ControlNets: Stability AI released three ControlNets for Stable Diffusion 3.5 Large. They guide composition from inputs such as edges or depth, giving designers far tighter control over layout than a text prompt alone.
• Platform API: Stable Diffusion models are available through the Stability Platform API using prepaid credits. Third-party guides report that one credit costs $0.01 and an SD 3.5 Large image uses 6.5 credits, with editing and upscaling services billed separately.
• Cloud partner access: Stable Diffusion 3.5 Large runs on Amazon Bedrock, Amazon SageMaker JumpStart and Microsoft Azure AI Foundry. Enterprises can add image generation inside cloud accounts they already govern, without managing GPUs themselves.
• Hardware optimisation: Stable Diffusion 3.5 has TensorRT builds for NVIDIA RTX GPUs and optimised versions for AMD Radeon GPUs and Ryzen AI chips. Stability AI says the NVIDIA builds run twice as fast with 40% less memory.
• Stable Assistant app: Stable Assistant gives non-technical users a browser interface to Stability’s image models, plus inpainting, upscaling and background removal. It runs on monthly credit plans with a three-day free trial.

Pros

• Free for most users: Stable Diffusion is free to self-host for individuals, researchers and businesses under $1 million in annual revenue. Hobbyists and small studios can generate unlimited images on their own hardware without per-image fees.
• You own the outputs: Stable Diffusion outputs belong to you under the Community License, including images from fine-tuned derivatives. Stability AI says you can use them at your discretion, which simplifies commercial client work.
• Deep customisation: Stable Diffusion’s open weights allow fine-tuning, ControlNets and custom pipelines that closed tools cannot match. Studios can train a model on their own style or product shots and keep it private.
• Runs offline and privately: Stable Diffusion can run entirely on a local GPU with no cloud connection. Prompts, reference images and results never leave your machine, which matters for confidential product or client material.
• Many ways to access: Stable Diffusion is reachable through local installs, the Stability API, AWS, Azure and Stable Assistant. You can start in a web app and move to self-hosting or an API as your volume grows.
• Model sizes for every budget: Stable Diffusion 3.5 comes in Large, Large Turbo and Medium versions with different hardware needs. You can match the model to a laptop GPU, a workstation or a cloud endpoint instead of paying for the biggest option.

Cons

• Steep local setup: Stable Diffusion needs a capable GPU, model downloads and a front-end tool before it runs locally. Beginners face a learning curve with drivers, VRAM limits and settings that hosted apps hide.
• Licence threshold for larger firms: Stable Diffusion requires an Enterprise Licence once commercial users pass $1 million in annual revenue. Growing companies must track that threshold and contact Stability AI, since enterprise prices are not published.
• Hardware demands: Stable Diffusion 3.5 Medium alone needs about 9.9 GB of VRAM before text encoders. Large models need more, so older or low-memory graphics cards will be slow or unable to run them.
• Confusing product line-up: Stable Diffusion now sits alongside Stable Image Ultra, Stable Assistant and Brand Studio, while DreamStudio redirects elsewhere. Newcomers can struggle to work out which product or plan actually gives them the model they want.
• Shifting company focus: Stability AI is moving towards commercial products such as Brand Studio, according to The Decoder’s April 2026 coverage. Open-model users cannot be sure how often new Stable Diffusion releases will arrive.
• Results need prompting skill: Stable Diffusion rewards careful prompts, negative prompts and settings tweaks to reach polished results. Casual users may find simpler hosted generators produce acceptable images with less trial and error.

Who is Using Stable Diffusion?

• Digital artists: Creating and fine-tuning custom styles.
• Marketing teams: Producing campaign and product visuals.
• Game studios: Concepting worlds, characters and assets.
• Developers and researchers: Building apps and studying image models.

Pricing

Stable Diffusion is free to self-host for most users, with paid routes through the API, Stable Assistant and enterprise licensing (official prices, October 2026).

API credit costs come from a June 2026 developer guide because Stability’s pricing console did not load for us, and that guide says Google sign-up adds 25 free credits.

Rates and credit allowances change often, so confirm them on Stability AI’s own pages before you budget for a project.

What Makes Stable Diffusion Unique?

Stable Diffusion gives you the actual model weights, not just access to a hosted generator. That means you can fine-tune it on your own images, add ControlNets, run it offline and keep full ownership of the results, all at no cost below the $1 million revenue line. Midjourney and DALL-E keep their models closed. The same models are also on AWS, Azure and Stability’s API, so a project can begin on a home GPU and scale to managed cloud hosting without switching model families.

Compatibilities and Integrations

• Self-hosting: Weights on Hugging Face with inference code on GitHub, for Windows, Linux and macOS setups.
• Cloud: Amazon Bedrock, SageMaker JumpStart and Microsoft Azure AI Foundry.
• API: Stability Platform API with prepaid credits for generation, editing and upscaling.
• Hardware: TensorRT builds for NVIDIA RTX and optimised builds for AMD Radeon and Ryzen AI.

Stable Diffusion Tutorials

Pick the access route that matches your hardware and budget.

1. Decide between Stable Assistant, the API or local self-hosting.
2. For self-hosting, check your GPU has enough VRAM for SD 3.5 Medium or Large.
3. Accept the Community License and download the weights from Hugging Face.
4. Load the model in your chosen inference tool and test short prompts.
5. Add ControlNets or fine-tuning once you have a reliable base workflow.

How We Rated It

Summary

Stable Diffusion suits artists, developers and studios who want control over an image model rather than a locked hosted service. Its main limitation is the technical effort and GPU memory needed to run it locally. Larger companies must also budget for an Enterprise Licence. Start with Stable Assistant’s free trial or SD 3.5 Medium on a modern GPU before committing to Large or paid API volume.