Run Stable Diffusion 3.5 your way.
Stable Diffusion 3.5 is Stability AI’s open image model family: three sizes released with downloadable weights, from an 8.1-billion-parameter professional model to a Medium build that runs on consumer hardware, under a licence that is free for most commercial use.
Capabilities and specifications on this page come from Stability AI’s official pages and published model documentation. Licensing terms and prices are set by the vendor and change without notice.
What Stable Diffusion 3.5 renders
These are Stability AI’s own examples from the Stable Diffusion 3.5 announcement and its image-models page. Together they cover what the family is sold on: photorealism, text inside the image, prompt adherence and style range.
Example images belong to Stability AI and are reproduced with credit: Stability AI image models
What Stable Diffusion 3.5 is
Stable Diffusion 3.5 is Stability AI’s open image model family, released in three sizes with downloadable weights. What makes it unusual here is the delivery: unlike the API-only models in this directory, SD3.5 can be self-hosted, fine-tuned and shipped commercially at no cost below a documented revenue threshold.
Three sizes, three jobs
Large is the 8.1-billion-parameter professional model; Large Turbo is the distilled build that reaches high quality in about four steps; Medium is the 2.5-billion-parameter one designed to run out of the box on consumer hardware.
Weights you can hold
All three variants are published on Hugging Face with the inference code on GitHub, so the model can be fine-tuned, optimised or run entirely offline.
A licence with a threshold
The Stability AI Community License is free for non-commercial use and for commercial use under $1M in annual revenue; above that threshold an enterprise licence applies, and outputs stay yours either way.
A trade-off the vendor admits
Stability states that variation between seeds on the same prompt is deliberate, preserving a broader knowledge base and diverse styles — and that unspecific prompts may therefore return more uncertain, aesthetically uneven results.
SD3.5 variants and the hosted services
Stability AI documents three downloadable SD3.5 variants plus two named hosted services. The rows summarise what the vendor says each one is for; the positioning is Stability’s own.
| Variant | What the vendor documents | Choose it for |
|---|---|---|
| Stable Diffusion 3.5 Large | 8.1 billion parameters; the most powerful model in the family, with the strongest prompt adherence. | Professional work at 1 megapixel. |
| Stable Diffusion 3.5 Large Turbo | A distilled build of Large that reaches high-quality output in about four steps, and is considerably faster. | Volume and iteration, where speed decides. |
| Stable Diffusion 3.5 Medium | 2.5 billion parameters on an improved MMDiT-X architecture, designed to run out of the box on consumer hardware. | Local generation between 0.25 and 2 megapixels. |
| Stable Image Ultra | A hosted service powered by the most advanced Stability models including SD3.5, aimed at photorealism. | Product imagery for marketing and advertising. |
| Stable Image Core | A hosted service built on an enhanced SDXL, tuned for speed at about half the price of SDXL. | High-volume generation where cost per image matters. |
Stability AI lists Stable Diffusion XL and SDXL Turbo as additional image models. Only the SD3.5 variants are open-weight releases; Stable Image Ultra and Core are API services with their own pricing.
What Stable Diffusion 3.5 does well
Every claim below is documented on Stability AI’s official pages. Figures the vendor does not publish — accuracy rates, head-to-head scores — are left out rather than estimated.
Text inside the image
Stability documents unprecedented text quality with fewer errors in spelling, kerning, letter forming and spacing, and shows it with legible lettering in the release examples.
Prompt adherence
The vendor’s own analysis places SD3.5 Large at the front of its class on following a prompt closely, and documents comprehension of long prompts that involve spatial reasoning, composition and style.
Versatile styles
3D, photography, painting, line art and effectively any visual style are all documented outputs, which is what makes one base model workable as the starting point for a fine-tune.
Diverse outputs
Stability documents images that represent a diversity of people and scenes without extensive prompting, and states that seed-to-seed variation is deliberate rather than a defect.
Customizability
The base models are built to be adapted: Query-Key Normalization in the transformer blocks stabilises training and simplifies downstream fine-tuning.
Documented specifications
Compiled from Stability AI’s official pages and the published model documentation. Where the vendor publishes no figure, that is stated instead of estimated.
- Developer
- Stability AI
- Open-weight variants
- SD3.5 Large · SD3.5 Large Turbo · SD3.5 Medium
- SD3.5 Large parameters
- 8.1 billion
- SD3.5 Medium parameters
- 2.5 billion
- SD3.5 Medium output range
- 0.25 to 2 megapixels
- SD3.5 Large output
- 1 megapixel
- SD3.5 Medium memory
- 9.9 GB VRAM, excluding text encoders, for full performance
- Large Turbo steps
- About four
- Licence
- Stability AI Community License
- Commercial use
- Free below $1M annual revenue; enterprise licence above it
- Output ownership
- The user retains ownership of generated media
- Distribution
- Weights on Hugging Face; inference code on GitHub
- Editing services
- Erase · inpaint · outpaint · conservative, creative and fast upscale · sketch, structure and style control
- Pricing
- Not reproduced here; Stability AI and its cloud partners set their own
Documented access channels
Stable Diffusion 3.5 is reachable in four materially different ways, from running the weights yourself to calling a partner cloud. They differ in cost and control, but not in licence terms.
Self-hosted weights
Download the models from Hugging Face and the inference code from GitHub, then run them on your own hardware.
OpenStability AI API
The official REST API exposes SD3.5 alongside Stable Image Ultra and Core and the editing, upscaling and control services.
OpenCloud partners
The same models are served through partner clouds, which removes the infrastructure work but not the licence terms.
Stable Assistant
Stability AI’s own web application exposes its models and editing tools with no local setup.
OpenStable Diffusion questions people search for
What is Stable Diffusion 3.5?
Stable Diffusion 3.5 is Stability AI’s open image model family, released in three sizes — Large, Large Turbo and Medium — with downloadable weights and a licence that is free for most commercial use.
Is Stable Diffusion 3.5 free?
For most users, yes. The Stability AI Community License is free for non-commercial use and for commercial use by organisations under $1M in annual revenue; above that an enterprise licence applies. Hosted services such as Stable Image Ultra are billed separately and are not covered by it.
Can I run Stable Diffusion 3.5 on my own computer?
Yes. The weights are published on Hugging Face. Stability documents SD3.5 Medium as running out of the box on consumer hardware and needing 9.9 GB of VRAM, excluding text encoders, to reach full performance; Large and Large Turbo need more.
What is the difference between the three SD3.5 variants?
Large is the 8.1-billion-parameter quality model, Large Turbo is a distilled version that reaches high quality in about four steps, and Medium is the 2.5-billion-parameter build designed for consumer hardware, generating between 0.25 and 2 megapixels.
Can Stable Diffusion 3.5 render readable text?
Yes — Stability documents unprecedented text quality with fewer spelling, kerning, letter-forming and spacing errors, and shows legible lettering in the official examples. Test it with your own copy anyway, because small type is where every image model eventually fails.
Is Stable Diffusion open source?
The SD3.5 weights are openly published and the licence permits commercial use below a revenue threshold. That is open weights under a community licence rather than an unrestricted open-source licence, and an organisation above the threshold needs a commercial licence.
Why do I get different images from the same prompt?
That variation is documented and intentional. Stability states that seed-to-seed variation preserves a broader knowledge base and diverse styles in the base models, and that prompts lacking specificity may therefore produce more uncertainty and uneven aesthetic results.