DiffusionBee
DiffusionBee is a free, open-source macOS app that runs Stable Diffusion entirely on your own Mac. It generates, edits, upscales and animates images from text prompts offline, with no account and no data sent to the cloud.
What is DiffusionBee?
DiffusionBee is a desktop application for macOS that runs Stable Diffusion directly on your own machine. It is a graphical front end, packaged with Electron, sitting on top of the open image models, so everything it does happens locally: no server call, no queue, no credits. The homepage sells it as 'The Ultimate Suite of Creative AI Tools' and as the fastest and easiest toolbox for running AI apps locally with Stable Diffusion.
The suite is built around the nine tools shown on the homepage: Text to Image, Generative Fill, Video Tools, Image To Image, Image Upscaler, Image Variants, Train Models, Control Images and Illusion Generator. An AI canvas sits alongside them, presented as a co-pilot that mixes prompt-driven generation with your own drawing. The model families announced in the GitHub README cover SD 1.x, SD 2.x, SD XL, inpainting models, ControlNet and LoRA, and the models themselves are downloaded from inside the application.
The product is completely free of charge, with no paid tier, no subscription and no in-app purchase, and its source code is published on GitHub under AGPL-3.0. The repository, divamgupta/diffusionbee-stable-diffusion-ui, was created in September 2022 and has gathered 13,580 stars and 727 forks; the app has been covered by TechCrunch, Fast Company, designboom, Digital Trends, AppleInsider and OSXDaily, and its community lives on a Discord server.
Distribution is deliberately plain: two .dmg installers hosted on GitHub Releases, one for Apple Silicon and one for 64-bit Intel Macs, with a one-click installer and no dependency to configure. That simplicity is also the boundary of the product. There is no Windows build, only a waiting-list form, and no Linux version, no web app, no mobile app and no API. No company is declared behind it: the footer reads '© 2024 DiffusionBee. All Rights Reserved.', the repository belongs to Divam Gupta, and neither a legal entity nor a contact address appears anywhere. The latest release, 2.5.3, dates from August 2024, while the download page still serves 2.5.1.
What it does
- Generate images from a text prompt, entirely on your own Mac
- Transform an existing picture with a prompt (image to image) and produce several variants of it
- Add or remove objects inside a selected region with generative fill, and extend an image beyond its frame with outpainting
- Upscale images to make them sharper and higher in resolution
- Generate animations and videos with the Video Tools, and optical illusions with the Illusion Generator
- Impose a specific composition on an image using control images (ControlNet)
- Train a custom model on your own dataset, or import a .ckpt model downloaded from Hugging Face
When to use DiffusionBee / When not to
A quick filter to help you decide if DiffusionBee is the right fit.
When to use DiffusionBee
- Mac owners on Apple Silicon who want unlimited AI image generation without a subscription, a quota or a credit card
- Privacy-conscious creators: prompts, models and generated images never leave the machine in normal use
- People who work offline or on locked-down machines, since no account, no API key and no connection are needed once the models are downloaded
- Newcomers to generative AI, thanks to a one-click installer that needs no dependencies and no technical knowledge
- Tinkerers who want to import custom .ckpt models from Hugging Face or train a model on their own pictures, entirely locally
When not to use DiffusionBee
- Windows and Linux users: only macOS builds exist, and Windows is limited to a waiting-list form
- Teams that need to work together, as there is no shared workspace, no account and no collaborative feature of any kind
- Developers who need programmatic access: DiffusionBee ships no API and no API documentation
- Owners of older Intel Macs without a dedicated GPU, where the vendor's own FAQ describes generation as very slow
- Organisations that require written commitments, since there is no privacy policy, no vendor terms of service and no support contact
How to use DiffusionBee
A typical end-to-end flow, from setup to results.
- Download the .dmg that matches your Mac from the download page: MacOS - Apple Silicon or MacOS - Intel 64 Bit
- Open the disk image and drag DiffusionBee into your Applications folder; there is no dependency to install
- Launch the app once and let it download and install the data it needs for generation
- Type a prompt in the Text to Image tab and click Generate
- Tune the settings if needed: number of images, height and width, steps, batch size, guidance scale and seed
- Reuse a seed (an integer from 0 to 4,294,967,295) with the same prompt and the same settings to reproduce exactly the same image
- Add a negative prompt to describe what the image should avoid
- Drop a PNG into Image to Image, or brush a mask over the area to repaint for inpainting; masking generously gives better results
- Move the 512x512 frame where you want to extend a picture and prompt again to outpaint, repeating in as many directions as you like
- Open the History tab to find earlier images with their prompts, settings and seeds, and add custom models through settings > add new model after downloading a .ckpt from Hugging Face
Pros & Cons
Pros
- Completely free, with no quota, no account and no credit card
- Privacy by construction: in normal use nothing leaves the device, prompts, models and images all stay on the Mac
- Works offline once the models have been downloaded
- One-click installer with no dependency to manage and no technical knowledge required
- A broad suite in a single app: generation, editing, upscaling, video, ControlNet and model training
- Open source under AGPL-3.0, so the local-only promise can be audited rather than merely believed
- Optimised for Apple Silicon, backed by 13,580 GitHub stars and coverage in TechCrunch, Fast Company, Digital Trends and AppleInsider
Cons
- macOS only: Windows is stuck on a waiting list and Linux does not exist
- Contradictory system requirements: the site FAQ asks for macOS 13.1 or later, while the GitHub README announces lower thresholds (Intel 12.3.1, M1 11.0.0)
- Speed depends entirely on the hardware: about 30 seconds per image on a MacBook Air M1 with 8 GB, and very slow on Intel machines without a dedicated GPU
- Resolution ceilings inherited from the models: inpainting capped at 512 px high and image-to-image outputs delivered in 512x512
- No legal documentation whatsoever: no privacy policy, no vendor terms of service, no legal notice and no published publisher identity
- No contact channel at all: no email, no form and no contact page, only a community Discord server
- Maintenance has stalled: no release since 14/08/2024, no commit since 30/10/2024, official documentation still written for version 1.5.1, a 'Help' link that returns a 404 and an /about page left on the default Jekyll template
Pricing & Plans
DiffusionBee is free in full. The site describes it as 'Completely free of charge', and there is no pricing page, no amount, no currency, no subscription and no in-app purchase anywhere on it — a complete archive enumeration of the domain confirms that a pricing page has never existed. There is therefore no entry price to quote, and no restricted free tier either, since the single edition that is distributed is the complete one. The only costs are indirect: a Mac capable of running the models, and the bandwidth needed to download the model weights on first use.
- Apple Silicon and MacOS - Intel 64 Bit
- are hardware variants of the same free build and not commercial tiers
- a third entry
- Windows 64 Bit
- only leads to a waiting-list form for a product that does not exist yet.
Data, GDPR & hosting
A consolidated view of how DiffusionBee handles your data.
GDPR overview
There is no GDPR mention at all: searching the whole site and the documentation returns zero occurrences of GDPR or of any equivalent wording. No privacy policy, no data protection officer, no Article 27 EU representative, no data processing agreement and no sub-processor list have been published. No company registration, no postal address and no contact email are disclosed either, so there is no identified controller to write to. What the architecture implies is more favourable than the paperwork: in normal use the processing is local, the publisher receives no personal data, and there is nothing to transfer, export or erase on its side. The optional sharing feature is the exception — it sends images, prompts and settings to a third party, arthub.ai, with no GDPR documentation on the DiffusionBee side. No compliance is claimed here, in either direction.
Who owns the data?
No privacy policy and no vendor terms of service exist — a full archive enumeration of the domain confirms those pages were never published — so no document formally assigns ownership of anything. The architecture answers in their place: DiffusionBee runs Stable Diffusion on your Mac and, in the vendor's own words, your prompts, models and generated images never leave your device. Generated images sit in ~/.diffusionbee/images/ and downloaded models in ~/.diffusionbee/downloads, on your own disk, under your own file permissions. The publisher receives nothing, so it has nothing to keep, sell or pass on to a third party. Deleting the ~/.diffusionbee/ folder removes the lot.
Reuse rights
The homepage FAQ states that you may use the images you generate, provided you respect the CreativeML Open RAIL-M model licence — that third-party licence, and not a DiffusionBee contract, is the only document the download is conditioned on, and it carries real use restrictions. No permission has to be asked of the publisher, who never sees the output in the first place. The vendor declares two exceptions to the local-only rule: downloading model weights, and voluntarily uploading an image. Sharing sends the image together with its prompt and its settings to arthub.ai, a third-party service that requires an account to be created. No telemetry is described anywhere — but no document rules it out either.
Data retention & training
Hosting summary
There is no vendor hosting to describe. DiffusionBee processes and stores everything on the user's own Mac: generated images go to ~/.diffusionbee/images/, downloaded models to ~/.diffusionbee/downloads, and the generation history lives in the same hidden folder. No hosting country and no hosting region is declared, and for normal use there is none to declare — the applicable jurisdiction is simply the one where the machine happens to sit. Two network exits are known: the download of the model weights, and the voluntary upload of an image to arthub.ai when the user explicitly chooses to share it, which is the only case where content leaves the device and lands on a third party's infrastructure. The showcase website itself is served from the United States behind Cloudflare (AS13335, anycast, resolved IP 172.67.203.191), but that concerns the web pages only: no prompt and no generated image ever transits through it.
Things to keep in mind
Risks and trade-offs to weigh before adopting DiffusionBee.
- The fully offline claim only holds after installation: the app still has to download model weights over the internet, and sharing an image sends that image, its prompt and its settings to arthub.ai, a third-party service that requires an account
- The images you generate fall under the CreativeML Open RAIL-M licence, which forbids a list of uses; read it before publishing, selling or commissioning anything made with the app
- The system requirements contradict each other, macOS 13.1 in the site FAQ against Intel 12.3.1 and M1 11.0.0 in the GitHub README, so verify on your own machine before relying on it
- The download page still serves 2.5.1 while the latest release is 2.5.3 of 14/08/2024, and no commit has been pushed since 30/10/2024: expect no security fix and no adaptation to future versions of macOS
- The official documentation is written for version 1.5.1 and warns that parts of it may not apply to other versions, the site's 'Help' link returns a 404 and the /about page was left on the default Jekyll demonstration text
- With no privacy policy, no vendor terms, no published identity and no contact address, you have no recourse whatsoever if something goes wrong: the only channel is a community Discord
- Unlimited free generation removes every natural brake, and it is easy to burn hours producing hundreds of variants instead of deciding what you actually want; meanwhile every image, prompt and setting piles up in ~/.diffusionbee/ on the machine, readable by anyone with access to your session, until you delete it yourself
Setup & Integrations
Technical difficulty
Very low. Download the .dmg, open it, drag the app across: the README promises a one-click installer with no dependencies and no technical knowledge needed. On first launch the app downloads the data required for generation by itself. There is no API key, no account and no command line for everyday use. The only decision is which build matches your processor, Apple Silicon or Intel 64-bit. The real constraint is hardware rather than software: the chip and the memory decide how pleasant it is. Advanced uses, such as importing a .ckpt model or training your own, ask for more familiarity.
Deployment
Integrations
Supported languages
Behind DiffusionBee
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Where are the images generated?
How much does DiffusionBee cost?
How long does it take to generate one image?
What do I need to run it?
Is there a Windows or a Linux version?
Do I have to create an account?
Can I use the images I generate?
Can I add my own models?
Where are my images stored, and how do I uninstall everything?
Is there an API or a support desk?
Should you pick DiffusionBee?
DiffusionBee is one of the rare generative-AI products that asks nothing of you: no account, no subscription, no credits, no connection. You download a .dmg, you generate, and everything — prompts, models, images — stays on your Mac. For a solo creator on an Apple Silicon machine, that combination of a genuinely broad creative suite and zero cost is hard to match, and the AGPL-3.0 source code makes the local-only promise auditable instead of merely stated.
The reservations are just as clear. Nothing here is contractual: no privacy policy, no vendor terms of service, no legal notice, no company name, no address, no email. If something goes wrong there is nobody to write to, only a community Discord. The maintenance picture points the same way: the last release dates from August 2024, the last commit from October 2024, the official documentation is still written for version 1.5.1, and the download page serves 2.5.1 when 2.5.3 exists. The site's own 'Help' link returns a 404 and its /about page was never filled in — small signs, but consistent ones.
The natural audience is therefore the individual: the curious beginner, the designer who wants to iterate without a meter running, the user who would rather not send images to someone else's servers. Teams, regulated environments and anyone who needs commitments in writing should look elsewhere, because there is nothing to hold on to here. Taken for what it is — an excellent, free, entirely local toolbox, generously given and no longer actively maintained — it remains one of the easiest ways to put image generation on a Mac.
- Choosing a selection results in a full page refresh.
- Opens in a new window.