Before you start: browser and graphics card setup

Important to set up! Otherwise Chrome may use the weaker, integrated graphics.

1 Model

Preset

Model files (custom file or URL)

Pick a file or enter a URL for each part. Use either a single-file checkpoint, or a diffusion model with its text encoder and VAE. Files are copied into browser storage once and reused on later visits.

2 Describe the image

Negative prompt

×

Send diagnostics (never prompts or images)

Size and steps

Width Height Steps CFG Seed New seed each time

Start from a picture

Drop an image here or anywhere on the page. img2img keeps its layout; edit models (FLUX.2 Klein) follow an instruction such as "make it night".

Choose or drop an image

Use the image as Strength

LoRA

A LoRA must match the model family (for example an SDXL LoRA for an SDXL model). Files are cached like model files.

Advanced

Sampler Scheduler VAE tile size Prediction Max VRAM (GiB)

Flash attention Auto-fit (stream weights from disk when VRAM looks tight)

Pick a model to start.

Your image will appear here made on your own computer

Open gallery

Recent

Log

Gallery

Every image you make is kept in this browser, on this device, with its prompt and settings. Downloaded PNGs carry the settings too: drop one back on the page to reuse them.

Stored only here Clearing site data deletes it, so export what you want to keep.

Benchmarks

How long one image takes, measured with the model already loaded. The first image after loading also compiles GPU shaders and takes longer.

Test your GPU

Runs SD-Turbo at 512×512 (1.9 GB download the first time): one warm-up image, then three timed ones. Images are not saved to the gallery.

Your results

Measured times

Reference machine: laptop with an RTX 5090 Laptop GPU; phone: Galaxy S23 Ultra. "—" means not measured yet. Times vary with power settings and other programs using the GPU.

More than a browser tab

The web app is free and stays free. Native apps for Windows and Android are in the works: same engine, more speed and bigger models.

Web

Free · available now

Runs in Chrome or Edge, nothing to install
Models up to about 10 GB
Gallery stored in your browser

Start creating

Desktop (Windows)

Coming soon · one-time purchase

Native GPU access, without the browser's memory limits
Bigger, higher-quality models such as Krea 2
Your own model and LoRA folders, files on disk
Works offline, no account

Android

In development

Uses the phone's NPU on Snapdragon 8 Gen 2 and newer
About 4 s per SD-Turbo image on a Galaxy S23 Ultra
Fully offline

"I want this" only counts a vote (no email, no account). It helps decide what ships first.

Support the free web app

The web app has no ads from ad networks and no trackers. If it saves you time, a coffee helps pay for the server and the GPUs it's tested on.

Buy me a coffee

Thank you! ♥ Coffee + Stable Diffusion = happiness

Help and setup

Important to set up! Otherwise Chrome may use the weaker, integrated graphics.

Browser and Windows setup

Storage on this device

Downloaded models

Privacy

Prompts and images are processed and stored only in this browser. Model files are downloaded from Hugging Face. If "Send diagnostics" is on (under the Generate button), the page sends browser and GPU details, timings and logs with prompts removed, to help fix failures.

Brulion

Image generation on your own graphics card, in the browser.

Powered by stable-diffusion.cpp + ggml

Models

Model files are downloaded from Hugging Face and come under their own licenses, named in each preset's description. Some of them don't allow commercial use.

Open-source licenses

This app includes the following open-source software. Their copyright notices and license terms:

All notices as a text file