Stable Diffusion · images
Stable Diffusion integration in a custom app
Stable Diffusion integration means your app sends a brief to an SD runner — hosted or your own GPU — and gets images back onto the case. Choose SD when you want to control model, LoRA, style and whether data leaves the building. Choose [DALL·E](/articles/integrations/dall-e-integration) when you want an API and to move on. What costs is not “a call”. It is queue, GPU, timeout, storage and the UI that lets a human pick among four proposals. Do not underestimate the UI. SMEs should almost never self-host in the first version. Hosted endpoint first. Self-host when you have volume and a requirement. Minimisation also applies to image briefs: do not send the customer’s face if a silhouette is enough.
Keep SD behind the same engine layer as OpenAI, so the image button is not a separate product. An AI engine, not an AI page.
What actually costs
- Queue and GPU time, not “an API call”
- Timeout when the runner is busy
- Storage of images and versions
- UI to pick among 4 proposals and store the prompt
Prompt
Menneske godkender før udsendelse
Hosted endpoint
- First version, low volume
- No GPU ops in-house
- You want style, not a server room
Own GPU
- A requirement that data stays
- Volume that justifies operations
- LoRA and weights you want to own
What to do now
Say whether you need to control model and data, or whether you just need images in the workflow. The first is SD. The second is more often DALL·E. Align that with us.
FAQ about Stable Diffusion integration
Should SMEs self-host?
Almost never in the first version. Hosted endpoint first. Self-host when you have volume and a requirement.
What is LoRA in this context?
A style adaptation so you do not start from scratch every time. It is a plus — and extra operations.
Is SD cheaper than DALL·E?
Per image, hosted SD can be cheaper at volume. GPU operations and UI eat the advantage if you run ten images a week.
May customer images be used as training?
Not without an agreement. Default: brief and output live on the case, not as training data.