jp.keijiro.ml-stable-diffusion

Unity Core ML Stable Diffusion Plugin
=====================================

!gif

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.

[Apple's Core ML port]: https://github.com/apple/ml-stable-diffusion

System Requirements
-------------------

- 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.

How To Try
----------

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.

[pre-converted Stable Diffusion 2 model repository]:
https://huggingface.co/apple/coreml-stable-diffusion-2-base

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.

Performance Considerations
--------------------------

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.

LCM (SD-Turbo) Support
----------------------

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.

[SD-Turbo]: https://huggingface.co/stabilityai/sd-turbo
[pre-converted SD-Turbo model]:
https://huggingface.co/keijiro-tk/coreml-sd-turbo

Sample Projects
---------------

!gif

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