Stable Diffusion

Where it sits

Stable Diffusion is the model family that started the entire text-to-image wave. When the first weights leaked and then shipped in 2022, it put generative imaging in the hands of anyone with a graphics card; four years of community building later, the surrounding tooling is arguably a bigger asset than the models themselves. In 2026 the family occupies a different position than its fame suggests: it is the open-weights option, the one you can download, host, fine-tune and ship inside your own product. Stability AI (London) released the current generation — the Stable Diffusion 3.5 suite — in October 2024: a Large model at 8 billion parameters, a Medium family at 2.5B/2B, and a Large Turbo variant for speed. Earlier generations (SD 1.5, SDXL) remain in wide circulation and still power a large share of the community tooling. What I look for in this category has shifted over the years from "can it make a pretty picture" to "can a team own the whole stack," and Stable Diffusion is still the only major family where the answer can be yes.

One correction to the folklore up front: Stable Diffusion is not simply "fully open source" any more, if it ever was. The 3.5 models ship under the Stability AI Community License (last updated July 5, 2024), which allows research, non-commercial and commercial use free of charge only for individuals and organizations under US$1,000,000 in annual revenue, requires commercial users to register with Stability, and terminates automatically above that revenue line, where an enterprise license is required. The older Core Models (including SD 3 2B) are free for everyone except commercial use above the same threshold. That is a permissive source-available license with a revenue tripwire — a meaningful distinction for anyone building a business on it.

What it actually does

There are three ways to run it. Self-host: download the weights from Hugging Face and run them through ComfyUI, the legacy Automatic1111 web UI, Forge, or your own pipeline. This is where the ecosystem is deepest — LoRAs, ControlNets, IP-Adapter, regional prompting, upscalers — and where your data never leaves your hardware. Hosted API: Stability's developer platform bills in credits at 1 credit = US$0.01, with 25 free credits on signup. At time of writing (platform.stability.ai/pricing), Stable Diffusion 3.5 Large runs 6.5 credits (US$0.065) per generation, Large Turbo 4 credits (US$0.04), Medium 3.5 credits (US$0.035), Stable Image Core 3 credits (US$0.03), Stable Image Ultra 8 credits (US$0.08), SD 3.5 Flash 2.5 credits (US$0.025), and SDXL from 0.9 credits. Editing endpoints (inpaint, outpaint, search-and-replace, search-and-recolor, remove background, relight), upscalers (US$0.02 to US$0.60 per image), control services (structure, sketch, style guide), Stable Audio, and 3D (Stable Fast 3D at 10 credits, SPAR3D at 4) all sit on the same credit wallet. The API allows 150 requests per 10 seconds and does not charge for outputs blocked by the content filter. The API surface also carries 3D and audio generation (Stable Audio 3.0 at 26 credits per generation), which matters if you want one vendor bill across modalities. Third, the hosted web app: DreamStudio has been rebranded Brand Studio, now subscription-shaped — a free one-time 500-credit trial, Creator at US$19/month with 2,000 credits, Core at US$50/month with 6,000, Enterprise custom.

My testing notes

Two observations from my testing. First, the self-hosted route is real but the hardware tax is too. Running SD 3.5 Large (8B) on a 24 GB consumer card worked, but iteration speed was slow enough that I reached for Medium or Large Turbo for concepting and reserved Large for final passes — and SDXL remains faster and, for some styles, better supported by community LoRAs than the 3.5 generation. If you do not already own a decent GPU, the "free" model costs more than a Midjourney subscription once you price in cloud compute or a new card.

Second, the hosted API is the most transparent billing in this list. I ran a 100-image batch through Stable Image Core at 3 credits per image (US$0.03), and the total matched the per-call rate exactly — flat per-generation pricing regardless of aspect ratio, no input-token trickery, and the 25 free signup credits were enough to validate my prompt set before spending anything. The 150-requests-per-10-seconds ceiling meant my client-side throttle, not the server, was the bottleneck. A third observation: the hosted and self-hosted experiences differ enough to count as different products. Brand Studio gives you the curated Stable Image services and a plain browser UI; self-hosting gives you the raw 3.5 checkpoints, community fine-tunes and every control mechanism the ecosystem has invented. My one irritation: credits issued from September 30, 2026 expire twelve months after purchase under the current terms, so buying a big pile in advance is no longer a store-of-value.

Who it is actually for

In practice I see three stable buyer profiles. Solo power users building a personal workflow in ComfyUI, for whom the cost is hardware and time. Product teams embedding generation behind their own interface, for whom the API's flat per-image pricing and the weights' license make unit economics predictable. And research groups, who use the family as a baseline because every paper and every community technique is written against it first. If you fit none of those profiles — you want a nice image for a slide deck — this is the wrong shelf.

Where it falls short

Out-of-the-box aesthetics are the standing weakness. Stable Diffusion 3.5's defaults are flatter and less "finished" looking than Midjourney or FLUX output; it needs prompt craft, a workflow (ComfyUI node graphs are a skill of their own) or fine-tuned checkpoints to reach competitor quality. The ecosystem fragmentation is real — Automatic1111, ComfyUI, Forge and a dozen forks each with their own extension quirks. The license tripwire (US$1M revenue, registration requirement, termination clause) is a genuine business constraint that "open source" glosses over. Stability AI itself has had well-documented management and financial turbulence — executive departures and restructuring through 2024, referenced in industry profiles of its alumni — which matters when your build depends on their hosted API or future model releases. For teams, there is also the compliance question the license creates: if your company crosses US$1M in total revenue — not revenue from Stable Diffusion output, total revenue — the free license terminates and you are in enterprise negotiation territory. I have watched startups miss this clause because it is easy to read "free for commercial use" and stop there. And there is no slick team workspace, no brand kit, no consumer polish: Brand Studio is a thin front end compared to Leonardo or Firefly.

Verdict

Rated 4.2 on the ToolSpotter rubric: core functionality depth is enormous when you count the self-hosted ecosystem — every control, adapter and fine-tune the community has built (the 40%); value is the best in the category below the US$1M line, since the weights are free and API cents are cheap (the 20%); usability is the weak leg — this is a power-user stack (the 20%); the LoRA/ControlNet/ComfyUI ecosystem is the largest in AI imaging (the 10%); and reliability signals are mixed: transparent API billing and a clear license on one side, corporate turbulence and expiring credits on the other (the 10%). If you need control, privacy or unit economics at scale, nothing beats owning the weights — just read the license before you print the business cards. If you need a beautiful image in ten seconds, buy one of the polished front ends instead.

Sources

• https://platform.stability.ai/pricing
• https://stability.ai/community-license
• https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/LICENSE.md
• https://huggingface.co/ckpt/stable-diffusion-3.5-medium/blame/main/LICENSE.md
• http://stability.ai/license
• https://theproofdesk.com/stability-ai-review
• https://recatools.com/ai-directory/dreamstudio
• https://developer.puter.com/tutorials/stability-ai-api-pricing
• https://unifuncs.com/s/Cd9TyTkS
• https://probsee.com/p/black-forest-labs