Diffusers is a native Mac app to generate images from a text description of what you want. It uses state-of-the-art models contributed by the community to the Hugging Face Hub, optimized and converted to Core ML for maximum performance. It is the Mac counterpart to the Diffusers open source library.
Models are downloaded on demand, and then cached in your Mac's hard drive. You don't need credits for online services and won't experience long queues – just generate all the images you want and use them for fun or work. Privacy is guaranteed: your prompts and images are yours to use, and will never leave your computer (unless you choose to share them with your friends or colleagues).
Diffusers leverages the best technologies to offer you the most convenient way to access Diffusion models like Stable Diffusion:
- Native Swift app designed for macOS that will feel at home in your Dock.
- Core ML models optimized for CPU, GPU and Apple Neural Engine.
- Diffusers library from Hugging Face, ported to Swift for maximum performance.
Diffusers launches with a set of 5 models, downloaded from the Hugging Face Hub:
- Stable Diffusion 1.4. This is the original Stable Diffusion model that changed the landscape of AI image generation.
- Stable Diffusion 1.5. Same architecture as 1.4, but trained on additional images with a focus on aesthetics.
- Stable Diffusion 2. Improved model, heavily retrained on millions of additional images.
- Stable Diffusion 2.1. The last reference in the Stable Diffusion family. Works great with "negative prompts" (see below) for details.
- OFA small v0. This is a special so-called distilled model, half the size of the others. It runs faster and requires less RAM, try it out if you find generation slow!
Diffusers allows you to tweak the generation parameters to help you achieve the images you want. These are the settings you can control:
- Model to use. Each model has its own personality and style, so people like to use different ones depending on the subject. Feel free to explore!
- Prompt and (optionally) negative prompt. The prompt is the description of what you want, and the negative prompt is what you _don't want_. Use it to tweak a previous generation (by removing unwanted items), or to provide hints for the model. For example, many people like to use negative prompts such as "ugly, bad quality" to make the model try harder. Or consider excluding terms like "3d" or "realistic" if you're after particular drawing styles.
- Guidance scale. Indicates how much the image should resemble the prompt. Low values produce more varied results, while excessively high ones may result in image artifacts such as posterization. Values between 7 and 10 are usually good choices, but results depend on the model too. Try it out!
- Seed. This is a number that allows you to reproduce a previous generation. Use it like this: select a seed and write a prompt, then generate an image. Next, maybe add a negative prompt or tweak the prompt slightly, and see how the result changes. Rinse and repeat until you are satisfied, or select a new seed to start over.
- Steps. How many "diffusion" steps is the image generation going to take. More steps can provide slightly better quality, and less steps makes the process run faster. The default value (25 steps) produces great quality in a reasonable amount of time.
Diffusers was created by Hugging Face. We are on a mission to democratize good machine learning, one commit at a time. Diffusers is fully open source and has a permissive license that allows you to build on it to create your own products. Check our support URL for details!
Stay tuned for more features, additional models and iOS/iPadOS support!
Version 1.1 of Diffusers is about Quality, Performance and UI improvements.
- Generate images up to twice as fast! We benchmarked across different computers and came up with simple heuristics to choose the best hardware (CPU, GPU and Neural Engine) where your models run. Speed improvements depend on your hardware.
- You can override the default selection to experiment yourself.
UI / New Features
- Guidance Scale can be configured.
- Safety checker can be disabled.
- Easily reuse a random seed.
- Display downloaded status for models.
- Provide additional info on models and generation settings.
- Display configured settings even when controls are not expanded.
- Reveal model folder in Finder.
- Allow generations to be canceled.
- Improve model download method to prevent errors and double downloads.
- Improved error reporting all around.
Diffusers is open source, we welcome bug reports, suggestions or contributions at https://github.com/huggingface/swift-coreml-diffusers
Ratings and Reviews
Great No-Frills experience, feels like it's made by Apple
If you just want to put in text and get back a picture, without fiddling with any of that complicated "Other Stuff", this app is absoolutely perfect. It leverages the Neural Engine nicely, has some options for model choices (more on that later), and best of all, the UI feels like it's made by Cupertino.
My main gripe is that I can't import my own models yet, but that's in development right now, and when it's released on the app store in a few monts, I'll probably update this to 5 stars.
I like it but
Has great potential, hopefully soon add custom model support, maybe add a button that will auto convert any model we download from civitai to the format that runs with M1 mac?
The 9second speed is good, my friends 3070 takes 8 seconds
I have a m1 max 64 GB
It's good enough as toy
Apple Neural Engine has different format with other model. So it's hard to add new model or other pipline work.
So , it is good enough as a toy.
Data Not Collected
The developer does not collect any data from this app.
Privacy practices may vary, for example, based on the features you use or your age. Learn More
- Hugging Face, Inc.
- 1.4 MB
- Requires macOS 13.1 or later.
- Age Rating
- © 2023 Hugging Face