Diffusion Bee

A one-sentence summary

Diffusion Bee is a local Stable Diffusion graphical interface for macOS; users can carry out text-to-image, image-to-image, local retouching, image expansion, ControlNet, and LoRA tasks without the need to set up a complex operating environment.

Tool Introduction

Diffusion Bee was created by Divam Gupta, and its source code is available on GitHub. Its main goal is to enable users of Intel and Apple Silicon Macs to run generative image models with just one click. The application performs the processing tasks locally, which is ideal for those who do not want to send their prompts and assets to cloud services continuously.

The latest publicly available version is 2.5.3, which introduces Flux.1, external Textual Inversion integration, and the capability to block inappropriate content. Flux.1 is available only on arm64 devices and requires macOS 13 or a higher version.

Overview of Core Competencies

Main functions

Text-to-image generation

After entering a natural language prompt, the application uses a local model to generate one or more images. Users can set the size, number of iterations, sampler, seed value, and prompt weights; the weaker the hardware, the longer the waiting time is usually.

Image-to-image generation

Image-to-image generation uses an existing image as a structural or visual starting point, and then creates a new version based on the given prompts. If the degree of change is too high, the details of the original image are lost; if it is too low, no noticeable changes can be seen.

Local redrawing and image expansion

Masking is used for local redrawing to identify the areas that need to be replaced, while expanding the image adds content outside its original boundaries. Both methods are suitable for fixing the composition and adding or removing elements, but light effects, perspective, and character limbs still require manual inspection.

Advanced AI Canvas

The infinite canvas allows for the drawing, combination, and expansion of images; users can create content in larger work areas step by step. Complex projects should be saved in stages to prevent changes made at one time from affecting already finalized areas.

ControlNet

ControlNet can generate structures by utilizing control information such as edges, depth, and posture. The available control models depend on the version of the application and the resources that have been downloaded; not all community-developed models are compatible.

LoRA and Textual Inversion

The 2.5 series supports combining multiple LoRAs during generation; version 2.5.3 also includes external Textual Inversion embeddings. Before importing, it is necessary to verify the base model series, file format, trigger words, license, and recommended weights.

SDXL and Flux.1

The application supports SD 1.x, SD 2.x, and SDXL; version 2.5.3 adds support for the Flux.1 model. Flux.1 is available only on Apple Silicon devices with macOS 13 or later, and Intel Macs do not have access to the same capabilities.

Local model download and import

Users can download models within the application, or they can import supported models from Hugging Face. Large models require significant disk space, and the import process may involve copying files; therefore, sufficient storage capacity should be reserved.

Generate history and zoom in

The application retains the generation history and provides image zooming to facilitate the retrieval of parameters and outputs. History, models, and cache are stored on the local device, and users are responsible for backing them up and cleaning them up as needed.

Download installation tutorial

1. First, check the processor type, macOS version, and available disk space in the Mac System Information.
2. Download the installation package corresponding to Intel or Apple Silicon from the official release page.
3. Open the installation package and place the application in the Applications directory.
4. On the first startup, allow the system to perform security checks and verify the application source.
5. Download the default model within the application, and wait for the weights to be fully written to local storage.
6. Enter simple prompts to generate test images, and check the speed, memory usage, and output.
7. After confirming that everything is working properly, download SDXL, Flux, or other large models.

Guide to using text-to-image generation

1. Select the base model that matches the device and task.
2. Organize positive prompts using subject, environment, composition, lighting, and style.
3. The negative prompts describe the defects that need to be reduced or the elements that are not required.
4. Start by testing with a smaller size and moderate step count; don’t aim for the largest possible screen from the beginning.
5. Fix the seed and modify one variable at a time to facilitate assessing the impact of parameters.
6. After selecting a satisfactory version, increase the resolution, redraw specific areas, or zoom in.
7. When saving the final output, record the model, LoRA, seed, and prompt at the same time.

Image-to-image generation and local modification process

1. Import the reference image for which usage rights are available, while keeping the original unmodified version.
2. Choose image-to-image generation or partial redrawing to create an accurate mask for the task.
3. Set the intensity of changes; start by using a moderate value to test the extent to which the original image is retained.
4. The description should only address the objects, materials, colors, or background that need to be changed.
5. Generate multiple samples and check the edges, shadows, perspective, and character structure.
6. Repair the abnormal areas one by one to avoid having to redraw the entire image repeatedly.
7. Finally, it is enlarged and fine adjustments are made using regular image software.

Model and functionality are compatible.

Which users are it suitable for

• Mac creator: Hope to be able to run Stable Diffusion without using the command line.
• Users who are sensitive to privacy: prefer that prompts and regular generated content remain on the local device.
• Designer: Quickly create concept sketches, composition plans, and style variations.
• Illustration enthusiasts: Use LoRA, ControlNet, and local redraw to control their works.
• Content team: Creates covers, illustrations, and draft visuals for social media offline.
• Developing learners: Read AGPL code and understand the structure of local inference applications.
• Apple Silicon users: Take advantage of inference paths optimized for M-series chips.
• Individual users who do not wish to bear the high costs associated with pay-per-use pricing in the cloud.

Typical use cases

• Generate article covers and concept illustrations based on the text description.
• Convert line drawings or sketches into various material and lighting schemes.
• Replace the background, objects, or specific clothing parts with local redrawing.
• Adjust vertical images to horizontal formats by enlarging them, or increase the amount of empty space in the image.
• Use ControlNet to reference pose, edges, or depth to maintain composition.
• Combine LoRA to create a specific style and save the reproducible parameters.
• Continue to use the downloaded local model in the absence of a stable network.
• After exploring design directions in bulk, move on to refining them using specialized software.

Product advantages

• One-click installation reduces the need for Python environments, dependencies, and command-line configurations.
• The main reasoning is carried out locally, and control over the materials is greater than in pure cloud-based services.
• It supports both Intel and Apple Silicon Macs.
• It covers text-to-image, image-to-image, redrawing, image expansion, zooming, and history.
• Compatible with SD 1.x, SD 2.x, SDXL, and Flux.1 on select hardware.
• It supports ControlNet, multiple LoRAs, and external Textual Inversion.
• The application is available for free use, and the number of generations generated is not charged based on cloud credits.
• The source code is made public and licensed under AGPL-3.0.

Usage restrictions and precautions

• Only a macOS version is available; there are no official applications for Windows, Linux, iOS, or Android.
• The local generation speed and available sizes are directly affected by the chip, memory, and temperature.
• Flux.1 is only supported on Apple Silicon and macOS 13 or later.
• The latest installation package for Intel Mac may require a higher system version than what is specified in older documents.
• Large models can occupy several GBs or even more of storage space, and importing them may create copies as well.
• Some SDXL, LoRA, or custom models may fail to import, result in black images, or be incompatible.
• The generated history and local files are not automatically backed up to the cloud.
• Running locally does not mean being offline forever; data is still transferred when the model is downloaded for the first time and when images are uploaded manually.
• AI may generate distorted figures, incorrect text, biases, or inappropriate content.
• The commercial rights depend on all the licenses related to the application, the base model, LoRA, reference images, and the intended use of the output.

Price and version

As of August 22, 2026, the Diffusion Bee application and its open-source code are available for free, with no official subscription plans or pricing based on the number of images processed. The main costs for users stem from the Mac hardware, electricity, storage space, and the models that need to be downloaded separately.

System requirements and platform

Privacy and local data

The project is designed to operate locally; aside from downloading model weights or when users upload images manually, the generated data is not sent to the cloud. Users should still check the network requests, the sources of the models, and any optional online features depending on the specific version in use.

• Prompt words, models, history, and outputs are mainly stored on the user’s Mac.
• The first time the model is downloaded, an internet connection is required, as it will connect to the model hosting service.
• When the upload function is used actively, the selected image leaves the device.
• Third-party model download sites have their own accounts, logs, and privacy policies.
• When multiple people use a Mac, separate system accounts and disk encryption should be employed.
• After sensitive images are generated, it is necessary to clean up the history, export the directories, clear the cache, and create backups.
• Before uninstalling the application, verify the models and historical directories to avoid accidentally deleting necessary works.
• Companies should evaluate open-source dependencies, the model supply chain, and local access permissions when using such technologies.

License and Commercial Usage Notes

The Diffusion Bee application repository is licensed under AGPL-3.0, but the models used for generation have their own separate licenses. The repository also reminds that outputs related to Stable Diffusion must comply with the CreativeML OpenRAIL-M terms.

APIs, SDKs, and open-source status

Diffusion Bee is an open-source desktop application; it is not a commercial API platform that offers cloud-based access. The developers have not released any public APIs for remote use, a key management console, or separate SDKs.

GitHub and version status

The official repository makes the application code, backend code, documentation, and Electron interface code available; the current latest version is 2.5.3. This version introduces Flux.1, external Textual Inversion, blocking of inappropriate content, and organization of the model pages.

The warehouse still has unresolved issues related to model downloading, importing, blacklists, and system compatibility. Before using community models, it is necessary to review the version notes and test them using copies; the list provided in the README should not be construed as indicating that all components are compatible without any exceptions.

Basic information

Recommendation score

The recommendation score is 4.5 out of 5. Diffusion Bee is easy to install, operates locally, and covers the main Stable Diffusion workflows; it is very attractive to users who wish to start using AI for image generation on a Mac as quickly as possible.

The main shortcomings are that it only supports macOS, the models require a large amount of storage space, and there are compatibility issues with some custom models. Users who need the latest Flux capabilities must use Apple Silicon and macOS 13 or later.

Frequently Asked Questions

Is Diffusion Bee free?

It is free; there is no official subscription or fee based on the number of images. Users still have to cover the costs of the Mac hardware, electricity, storage, and any third-party models that may be needed.

Is it necessary to register an account?

For ordinary local generation, no Diffusion Bee account is required. When downloading certain third-party models, the platform where those models are hosted may demand an account or consent under a license.

Is it compatible with Windows?

It is not supported. The official application is intended for macOS users; Windows and Linux users should choose other trusted Stable Diffusion tools.

Are Chinese prompts supported?

Chinese can be entered in the interface, but the quality of comprehension depends on the training data used for the specific model. Many models perform more stably when given prompts in English, and this can be tested by making comparisons.

Will the data be uploaded to the cloud?

Normal generation is primarily carried out on the local device. Downloading models requires an internet connection, and when users upload images, the corresponding files leave the local device as well; therefore, it cannot be considered a completely offline process.

Is Flux.1 supported?

2.5.3 Supports Flux.1, but it is only available for Apple Silicon and requires macOS 13 or later. The actual speed and available sizes also depend on the amount of memory.

Can LoRA be imported?

Yes, the 2.5 series supports the use of multiple LoRAs during generation. It is necessary to ensure that the LoRAs are compatible with the base model series before importing them.

Can the generated images be used for commercial purposes?

Just because an application is free does not mean it is permissible to use it; it is necessary to check the licenses and rights related to the underlying model, LoRA, embeddings, reference images, and the intended purpose of the output as well.

Are APIs provided?

There are no public cloud APIs or official SDKs. It is a desktop application that runs locally, and its open-source code is suitable for research and customization.

Is Diffusion Bee open source?

It is open source; the official repository uses AGPL-3.0. The model weights are covered by a separate license, and this license cannot be conflated with the license of the application code.

Summary

Diffusion Bee reduces the barriers to using local AI-based drawing tools on Mac, offering in one interface text generation, modification of reference images, local repair, structural control, and model expansion.

Before making a selection, it is necessary to verify the chip and system version, reserve space for the models, and examine each model’s compatibility as well as its commercial licensing terms. Users working with sensitive material should also pay attention to the network boundaries related to model downloading and uploading.