pAIring image generation

pAIring can create images on the Images page. It uses stable-diffusion.cpp, a light engine that runs on a plain CPU (no GPU or PyTorch needed) and can also use a GPU.

Set it up (admin)

Open Settings → External servers and choose Image-generation engine.
Under stable-diffusion.cpp, pick Local and press Install from GitHub. pAIring downloads a ready-made build. Choose the build for your hardware:
- Auto (GPU if available)
- CPU only
- GPU, Vulkan (NVIDIA, AMD or Intel)
- GPU, ROCm (AMD)
Or, on Linux, press Build from source. This needs git, cmake and g++:
Terminal window
```
sudo apt install git cmake build-essential
```
The source stays on disk, so later rebuilds only recompile what changed. It can take a while on an old CPU.
Download an image model. Open Settings → Models → Browse more models and search, for example, sd-turbo. Image models are detected and installed into the image engine’s own folder, not into Ollama or Matricxon. Some models need extra files (a text encoder or a VAE). pAIring downloads those for you and passes them to the engine.
Back in the image engine section, choose the Model and press Start.

The engine serves one model at a time. To switch models, choose another in the Model list and start again.

You can also use Remote mode to point at one or more sd-server nodes (up to 120). Then the models are managed on those nodes.

Engine options

Option	What it does
Mode	Local (pAIring manages it) or Remote (external servers)
Build	Auto, CPU only, Vulkan or ROCm
Version and repo	Use another release or fork instead of the default tested one
sd-server path	For an engine you installed yourself
Models folder	Where image models are stored
Extra launch args	Extra command-line options for the engine
Generation timeout	Minutes before a job is given up as failed
Images folder	Where finished images are saved

Create an image

Open Images (in the sidebar). Fill in the form and press Generate.

Setting	What it does	Range	Default
Mode	Text to image, or Image to image		Text to image
Prompt	Describe what you want	up to 4000 characters
Negative prompt	Describe what to avoid (optional)	up to 4000 characters
Model	Which image model to use
Width, Height	Size in pixels	64 to 2048	512
Steps	More steps can add detail but take longer	1 to 150	1
CFG	How strictly the image follows the prompt	0 to 30	1
Seed	A fixed number repeats the same image. Random picks a new one	any	Random

The defaults of 1 step and CFG 1 suit fast models such as SD-Turbo. Other models usually want more steps and a higher CFG, so check the model’s page.

Image to image

Choose Image to image, add a source image, and set Strength:

Strength goes from just above 0 to 1 (0.75 by default). A low value stays close to your picture. A high value changes it more.
This mode needs the stable-diffusion.cpp engine.
Your source picture is kept in memory only while the job starts. It is never saved to disk or the database.
If the server restarts before the job begins, the picture is gone and you need to submit it again.

While it runs and afterwards

A running job shows a progress bar, the time elapsed and a rough time left. Press Cancel to stop a queued or running job.
You can leave the page. The job keeps running and you can come back to it.
Your generations appear as tiles, newest first, up to your last 100.
Save downloads an image. The × deletes it. Deleting a finished image asks you to confirm, and it cannot be undone.
Your images are private. Other users cannot see or open them.
A job interrupted by a server restart is marked as an error.

Tips

Start with a small size (512) and few steps, and raise them once you like the result.
A CPU-only node can take many minutes for one image, so set the timeout high enough.