Stable Diffusion plugin for Unity, based on Apple's Core ML port. You can run the model on-editor and at-runtime without needing any extra components.

• Unity 2023.1 or later
• Apple Silicon Mac (editor/runtime support) with macOS 13.1 or later
• iPad Pro with Apple silicon (runtime support) with iOS 16.2 or later

Although the plugin supports iOS, it requires huge amount of memory to run the model, so it only supports memory-rich iPad models.

Before running the sample project, you must put the model files in the Assets/StreamingAssets directory.

• Clone or download the pre-converted Stable Diffusion 2 model repository.
• Copy the split_einsum/compiled directory into Assets/StreamingAssets.
• Rename the directory to StableDiffusion.

It takes a long time (a few minutes) for the first run. After this initialization step, it only takes a few tens of seconds to generate an image.

You can change which processing unit it uses by switching the "Compute Units" property in the Tester component.

It depends on the device model to choose the best option:

• M1/M2 Mac and iOS: GPUs in those devices are not powerful enough, so you should select "CPU and NE" (Neural Engine) or "All".
• M1/M2 Pro/Max Mac: GPUs in those devices have enough processing power compared to NE, so "CPU and GPU" can be a better option.

When using "CPU and GPU" mode, you must use the "original" model instead of the "split_einsum" model. Please overwrite the StreamingAssets/StableDiffusion directory with the original/compiled directory.

You can use SD-Turbo or other LCMs (latent consistency models) with setting Pipeline.Scheduler to Lcm. You might also have to change StepCount to 1~4 and GuidanceScale to 1~2. Please refer to the model description to know the correct settings.

You can download the pre-converted SD-Turbo model from my Hugging Face repository.

• Flipbook3: Running the image-to-image pipeline with a real-time 3D scene.