stable-diffusion.cpp
Tool, image video, by leejet
Assessment
The way to get a picture out of a machine that cannot run ComfyUI. Fine for quick images that decorate something; for anything people will look at closely, send the job to the big box.
2026-10-09
Strengths
- SD 1.5 images on hardware nothing else supports: a 4 GB AMD card through Vulkan, or a CPU
- One binary, a tiny footprint, a simple server
- Good enough for an illustration on a reminder or a quick concept
Limitations
- Slow: a 512-pixel image in about 35 seconds on a 4 GB card, with spill into system RAM
- SD 1.5 quality; SDXL and newer models want more memory than it has
stable-diffusion.cpp is a C/C++ implementation of Stable Diffusion inference in the ggml family, with Vulkan, CUDA, Metal and CPU backends.[1] It runs SD 1.5 class models in a couple of gigabytes, which puts image generation on cards and machines the Python stack cannot use.
What it is for
Quick images from a small machine: an illustration for a reminder, a placeholder, a rough concept to react to. It keeps a small box useful for images without pulling the big GPUs away from the language model or video work.
What it costs
On an RX 6500 XT with 4 GB, a 512-pixel image took about 35 seconds, during which VRAM went from 2.4 to 3.2 GB and about a gigabyte spilled into system RAM.[2] Live speech-to-text on the same card kept answering, which is why images stay on that card; a render is still half a minute of contention, so a job that retries failures without a cap can hurt a phone call.
Things to know
- Keep the model at SD 1.5 size on a 4 GB card; SDXL fills it and the spill makes everything slow.
- A switch to turn image generation off is worth having where the card also serves calls.
- For quality, ComfyUI on the big cards; this is the fallback and the quick path.
History
- 2026-10-09: article written and published.
- 2026-10-09: entry created.
Comments
Public comments on each entry are coming. Nothing is collected here yet.