iOS app - AR Diffusion Museum

An iOS app that generates images using Stable Diffusion and displays them in AR.

You can generate images specifying any prompt (text) and display them on the wall in AR.

macOS 13.1 or newer, Xcode 14.2 or newer
iPhone Pro 12+ / iOS 16.2+, iPad Pro with M1/M2 / iPadOS 16.2+

You can run the app on above mobile devices. And you can run the app on Mac, building as a Designed for iPad app.

This Xcode project uses the Apple/ml-stable-diffusion Swift Package.

Apple/ml-stable-diffusion Swift Package/Library
SwiftUI, ARKit, RealityKit

This project does not contain the CoreML models of Stable Diffusion v2 (SD2). You need to make them converting the PyTorch SD2 models using Apple converter tools. You can find the instructions of converting models in Apple's repo on GitHub.

Apple/ml-stable-diffusion repo: https://github.com/apple/ml-stable-diffusion

There is a Readme in another GitHub Repository that explains how to add Stable Diffusion CoreML models to your Xcode project. Please refer to it.

Image Generator with Stable Diffusion v2: https://github.com/ynagatomo/ImgGenSD2

Change Log

[1.3.0 (7)] - Feb 10, 2023 [Changed]
- changed generateImages(of:enableStableDiffusion:) function to adapt to apple/ml-stable-diffusion v0.2.0 API change on generateImages(configuration:progressHandler:)
- apple/ml-stable-diffusion v0.2.0 is required.
[1.2.1 (6)] - Jan 1, 2023 [Changed]
- changed the Guidance Scale range from 0...10 to 0...50
- changed the Seed range from 0...1000 to 0...UInt32.max and added a TextField for the Seed
[1.2.0 (5)] - Dec 25, 2022 [Added]
- added the Guidance Scale. It requires the latest apple/ml-stable-diffusion Swift Package.
[1.1.0 (4)] - Dec 21, 2022 [Added]
- added the Negative Prompt. It requires the latest apple/ml-stable-diffusion Swift Package.
[1.0.2 (3)] - Dec 12, 2022 [Merged]
- merged the PR by Sergey Dikarev. It uses the configuration, MLModelConfiguration.computeUnits = .cpuAndGPU, when creating a StableDiffusionPipeline. This configuration is suitable for mobile devices. And it adds two entitlements, Extended Virtual Addressing and Increased Memory Limit. They are suitable for an execution on iPhone.
- if you encounter signing and capability errors for the entitlements, make sure that you are using the App ID which are registered the capabilities, "Extended Virtual Address Space" and "Increased Memory Limit", at Developer - Identifiers site.
- not use the .cpuAndGPU config when running on macOS, otherwise it causes errors.
- disable displaying intermediate images when running on devices, to improve the performance.
[1.0.1 (2)] - 2022-12-16 [Changed]
- use apple/ml-stable-diffusion Swift Package v0.1.0.
- set reduceMemory option of StableDiffusionPipeline(resource:) to true.
- iPhone requirements was changed to iPhone Pro 12+.

Features

image generation using Stable Diffusion v2 on device
showing generating images step by step
saving generated images in Photo Library
displaying generated images on the wall in AR
automatic switching of displayed images at regular intervals
automatic enlargement according to viewing distance (Large projection on outdoor walls)
built-in sample images

UI

This project provides a minimal UI. Feel free to extend it as you like.

Consideration

MPS internal error

Currently, using CoreML Stable Diffusion Library and RealityKit API such as ModelEntity.load(name:) together often causes MPS internal errors.
As a workaround, the 3D model of the picture frame is replaced with a simple one now.

In action on iPad: image generation => AR display