picoDiffusion

A minimal AI image generator for Stable Diffusion 1.5.
No bloat. No complexity. Just you, a prompt, and a GPU.

picoDiffusion interface showing a generated image

Minimal and focused

One page, one purpose. Every control has a built-in help guide that explains what it does and how to use it. No hidden menus or 47 tabs of settings.

Built for learning

Every control has a help button that explains what it does and what values to try. This is a tool for learning how image generation works, not a production pipeline.

Docker

Runs in a Docker container with NVIDIA GPU passthrough. One command to build, one command to start. Works on any Linux system with Docker.

Fast-ish on modest hardware

Generates 512x512 images in under a minute on a Quadro P2000 with 4GB VRAM. No expensive GPU required.

LoRA support

Load LoRA add-ons to change styles or add concepts. Adjustable weight slider lets you control how much influence the LoRA has.

Lazy model loading*

No model is loaded at startup. Choose your checkpoint from the UI and it loads on demand. Switch models without restarting the container.

* This card is here to balance out the other cards

Flexible output sizes

Width and height sliders from 256 to 1024 in steps of 32. Make squares, portraits, landscapes — whatever your image needs.

PNG metadata

Every generated image has its full settings embedded in the PNG file. Prompt, seed, checkpoint, LoRA, sampler — all saved so you can reproduce any image.

What you need for the best experience and results

That is it. No Python environment to set up, no dependencies to manage, no CUDA toolkit to install. Docker handles everything.

Local use only. picoDiffusion is designed to run on your own computer for personal use. It has no user authentication, no access controls, and an open API. Do not expose it to the internet or use it to host image generation for others.
Get Started