🟢 All systems operational · Image models on hardware you rent
+ Add Provider
Open weights · Community License

Stable Diffusion —
Your Card, Your Images

Stability AI publishes the weights for its image models, so the whole pipeline can run on a machine you rent: no per-image fee, no queue, no content filter you did not write. The two things worth reading before you start are the licence and the memory figure, and both are below.

See what a GPU costs → Read the licence first
stable-diffusion — serve
$ pip install diffusers transformers accelerate
✓ weights load from Hugging Face
# or run a UI over the same weights
$ python main.py --listen
✓ node graph in the browser
# the card decides how long each image takes,
# the licence decides whether you may sell it

What self-hosting changes

You keep the images
Stability states that users retain ownership of the media they generate, without restrictive licensing on the output. That is a claim about the pictures, not about the model — the model has its own terms, one card to the right.
Not open source, and it matters
SD 3.5 ships under the Stability AI Community License: free for non-commercial use, free commercially only while the organisation is under $1M in annual revenue, and an Enterprise licence above that. Read the section below before you build a product on it.
Three sizes, one family
SD 3.5 Large is 8.1 billion parameters aimed at 1 megapixel. Large Turbo is a distilled version that produces an image in four steps. Medium is 2.5 billion, built for consumer hardware at 0.25 to 2 megapixels.
No per-image bill
A rented card costs the same whether it makes ten images a month or ten thousand. Somewhere between those two numbers a hosted API stops being cheaper, and where that point sits is arithmetic only you can do.
The ecosystem comes with it
LoRAs, ControlNets, fine-tunes and the node-graph interfaces all attach to weights you hold locally. None of that is available through somebody else’s endpoint.
Moderation is now yours
A hosted service refuses some prompts on your behalf. Your own box does not, which means the policy, the logging and the liability all move to you. That is a responsibility, not a feature.

The licence, in Stability’s own terms

“Open weights” is not the same as “open source”. Stability AI publishes SD 3.5 under the Stability AI Community License, and states the terms as: free for non-commercial use; free for commercial use up to $1M in annual revenue; above that threshold, an Enterprise licence obtained from Stability. It also states that you retain ownership of the media generated, without restrictive licensing on the output. The older SDXL 1.0 sits under a different licence again, CreativeML Open RAIL++-M. These are summaries of what the vendor publishes, not legal advice, and the revenue threshold is measured on your organisation, not on this project. Read the licence that ships with the exact checkpoint you download.

How much GPU memory each model needs

One published figure, the rest is our arithmetic. Stability states that SD 3.5 Medium needs 9.9 GB of VRAM, and says explicitly that this excludes the text encoders. For the others we do the sum in the open: weights take roughly 2 bytes per parameter at FP16, and the text encoders, the VAE and the working buffers sit on top. Diffusion models do not size like language models — there is no growing context here. What pushes the number up instead is resolution, batch size and whether you keep the encoders resident or swap them in and out.
Model Parameters Weights at FP16 Comfortable card Source of the figure
SDXL 1.0 base ≈ 3B ≈ 6 GB 8–12 GB Ours, from the parameter count
SD 3.5 Medium 2.5B ≈ 5 GB 12 GB 9.9 GB stated by Stability, excluding text encoders
SD 3.5 Large 8.1B ≈ 16 GB 24 GB Ours, from the parameter count
SD 3.5 Large Turbo 8.1B ≈ 16 GB 24 GB Same weights as Large, four steps instead of many

The gap between the two right-hand columns is the point: SD 3.5 Medium’s weights are about 5 GB and Stability still asks for 9.9 GB, because the text encoders and the working buffers are most of the bill. Anything above 24 GB on this page is for batching several images at once or for fine-tuning, not for generating one picture.

Cheapest GPU at each tier

All GPU plans →

The cheapest live GPU plan in the DohoHub catalogue at each memory tier, out of 142 plans whose provider publishes a GPU memory figure. Prices come from our own feed, normalised so they are comparable and shown in the currency picked in the header. Where one machine is the cheapest way to clear several tiers, it appears once and the note says which tiers it covers.

Provider Plan Specs Price Visit
Servarica
Bee GPU VPS
cheapest at 8–24 GB+ · 24 GB on board
 6 Cores24 GB vRAM30 GB RAM400 GB NVME100 TB traffic10 Gbps
$79.00/mo
 Visit
NovoServe
Supermicro X11 10SFF (GPU)
cheapest at 48 GB+ · 64 GB on board
 64 GB vRAM25 TB traffic
$355.99/mo
renews at $402.52
 Visit

Image generation is one of the few GPU workloads where the card’s generation matters more than its memory, because a picture is a fixed amount of compute rather than a stream of tokens. We record the memory a plan advertises and not the model of card, so a plan that clears the memory bar can still be several times slower per image than its neighbour. None of these providers sells “Stable Diffusion hosting”; they sell GPU servers.

Pick a card that fits the model you want

Filter the catalogue by GPU memory, cores, storage and location — live prices, tracked every six hours.

Find a GPU server →
↑