Easy Diffusion: the installer that puts Stable Diffusion behind a browser UI

Two download buttons, no package manager

Installation is a single file per platform from the releases page: `Easy-Diffusion-Windows.exe`, `Easy-Diffusion-Mac.zip` or `Easy-Diffusion-Linux.zip`. There is no pip step, no conda environment to create and no model to download before the first run. The README states the aim plainly, that it does not require technical knowledge and does not require pre-installed software.

On Windows you run the installer, then launch `Easy Diffusion` from the Start Menu or a desktop shortcut. One instruction is worth repeating because it is a known failure mode: if SmartScreen blocks the program, click `More info` and then `Run anyway`. The other is that on Windows 10 you should install at the top level of a drive, for example `C:\EasyDiffusion` or `D:\EasyDiffusion`, because of file path length limits. That advice exists because the bundled packages create deeply nested paths, and Windows path limits are not something the installer can work around for you.

On Linux and Mac you unzip the `easy-diffusion` folder, open a terminal in it, and start the launcher:

cd easy-diffusion
./start.sh

The README also accepts `bash start.sh` if the execute bit did not survive the unzip. Uninstalling is deleting the folder, since everything downloaded lives inside it.

The hardware floor is a 2 GB card, or a CPU, or both

The requirements section is specific enough to decide before you download. On Windows and Linux you need an NVIDIA or AMD graphics card with at least 2 GB of memory, or you run on your CPU. Two footnotes carry the detail: NVIDIA cards need CUDA compute capability 3.7 or higher, and AMD needs ROCm 5.2 or newer. On Mac the listed options are M1 through M4, or an AMD card in an Intel Mac, or the CPU again.

Above that, two system numbers apply everywhere: at least 8 GB of system RAM and at least 25 GB of free disk space. The disk figure is the one people are surprised by, because it covers the application, the bundled environment and the models you download later rather than the application alone.

The performance claim in the README is a single concrete sentence: a 512x512 image with euler_a in 5 seconds on an NVIDIA 3060 12GB. The memory claim is stated separately, that 512x512 images can be made with under 2 GB of GPU RAM and 768x768 with under 3 GB. Those are the project's figures for its own defaults, and they are the right baseline to compare a real card against. Anything below 2 GB of VRAM will still run, slowly, on the CPU path.

Enabling the newer engine is a settings change and a restart

Version 3.0.16, published 2026-03-31, added support for Flux 2 Klein, Z-Image, Qwen and quantized gguf files, along with better support for AMD and Intel GPUs including integrated ones. None of it is on by default. The release notes give the sequence:

1. Open the `Settings` tab in Easy Diffusion's UI. 2. Select `v4 engine` in the Backend option. 3. Press `Save` and restart Easy Diffusion.

The same pattern appeared in v3.0.13, published 2025-08-08, which added Chroma and Krea behind a `v3.5 engine` option that first required enabling `Beta` in Settings. So there is a pattern worth knowing: new models land on a versioned engine that you opt into, and switching is always Settings, Save, restart. There is no in-app migration to think about.

Note the mismatch between the README title, which still reads Easy Diffusion 3.0, and the engine versions now in play. The release line is v3.0.x while the selectable backend goes to v4, which is a naming inconsistency rather than a technical one, but it means you should read the release notes rather than the README heading to work out what is current.

The earlier v3.0.9c release from 2025-03-06 is the one that widened the base: it moved to Python 3.9 and added NVIDIA 50xx series, which is the Blackwell generation, alongside AMD GPUs and integrated GPUs on Windows.

A task queue, live preview and modifiers instead of a config file

The feature list is long, but it clusters into a few ideas. The Task Queue is the one that changes how you work: you can queue all your ideas without waiting for the current task to finish, which means a session is a list of things to try rather than one prompt at a time.

Live Preview shows the image while it is being drawn, which turns a wait into feedback. Image Modifiers are a library of style tags such as Realistic, Pencil Sketch and ArtStation that you can apply to experiment with looks quickly. Multiple Prompts File lets you queue a whole batch by putting one prompt per line in a text file. The searchable models dropdown lets you organise models into sub-folders and still find them.

Generation modes are Text to Image, Image to Image and Inpainting, with ControlNet for pose or outline control and a simple built-in drawing tool so you do not need an external program to sketch something for the model to fill in. There are sixteen named samplers, from PLMS and DDIM through DPM++ 2m SDE and UniPC, and prompt control is the standard Stable Diffusion syntax: `+` and `-` for attention, weighted terms like `(red)2.4 (dragon)1.2`, a pipe for a prompt matrix, and braces for a prompt set.

After generation, the post-processing is one click: RealESRGAN upscaling and GFPGAN face correction, plus a Loopback option that feeds the output image back in as the input for the next task. Output goes to disk as JPEG, PNG or WEBP.

Custom models, plugins and the licensing picture

The advanced half of the documentation is about replacing what ships with the bundle. You use your own `.ckpt` or `.safetensors` file by placing it inside the `models/stable-diffusion` folder, and from there you can merge models, use a custom VAE, supply your own GFPGAN models, and use Textual Inversion embeddings to push a particular concept hard.

There is a UI plugin system described as a growing list of community-generated plugins in the project wiki, with the invitation to write your own. That is the extension mechanism, and for a project whose whole pitch is a friendly interface it is the part most likely to change the ceiling on what the interface can do.

Licensing is where the repository does not simplify things. The project metadata records no standard SPDX identifier, and the tree contains a `LICENSE` file, a separate `CreativeML Open RAIL-M License` file, and a `3rd-PARTY-LICENSES` directory. The RAIL-M file is the licence that travels with Stable Diffusion weights and code, so the bundling carries that obligation with it. Anyone planning to redistribute a build should read all three rather than assuming a single permissive licence.

The repository also has `PRIVACY.md`, `CHANGES.md`, `CONTRIBUTING.md` and a `NSIS/` directory, which is the installer script directory for the Windows executable. `build.bat` and `build.sh` sit at the root, `environment.yaml` describes the bundled Python environment, and `package.json` contains only two Prettier scripts, which is a fair signal of how little of the application is actually JavaScript.

Where the README hands off to the wiki

The README is an advertisement and an install guide, and it knows it. Three links sit at the top of the page and they are where the real documentation lives: an installation guide anchor, a Troubleshooting guide in the wiki, and a User guide in the wiki. A Discord invite follows for support queries and development discussion.

That split is worth planning around. Installation, hardware, the feature list and the release instructions are all on the README page and are complete enough to get running. Everything about why a generation produced a bad image, how to structure workflows, and how the plugins work is wiki territory.

Compared with ComfyUI, which the related searches suggest readers are weighing it against, the trade is straightforward. ComfyUI exposes a node graph and every parameter, which is more power and a steeper first hour. Easy Diffusion gives you a form, a queue and a preview, and hides the sampling parameters behind a dropdown. The comparison also runs the other way on model support: ComfyUI's community tracks new architectures faster, while here you wait for a release and then switch engines in Settings.

The project is not abandoned. The last push was 2026-09-11, and the recorded releases show a pattern of small version bumps carrying model support rather than large rewrites. At 10463 stars, 864 forks and 96 open issues, it is also large enough that the wiki is a better description of the current state than the README is.

Editorial conclusion

Easy Diffusion earns its place for the first hour of using local image generation, because it removes the part where you assemble a Python environment, a model folder layout and a launch flag yourself. The `models/stable-diffusion` drop-in convention, the searchable models dropdown and the Intelligent Model Detection that picks a YAML config for a model you added are the details that would otherwise be your job. What it does not remove is the hardware floor: 25 GB of disk, 8 GB of system RAM, and either a card with at least 2 GB of memory or a CPU run. Version 3.0.16 turned on the newer engine by default path for Flux 2 Klein, Z-Image and quantized gguf files, so start there, drop your own safetensors into `models/stable-diffusion`, and move to ComfyUI only once the node graph becomes the thing you actually want.

Frequently asked questions

Official sources

1. easydiffusion/easydiffusion on GitHub
2. Issues
3. Project website
4. README
5. Releases