Pipeline reduceMemory = true Crash on iOS

[TimYao18]

I test the StableDiffusionPipeline API with Diffusers app on both macOS and iOS and try to let users select their own custom models into the app.

I put custom Stable-Diffusion-v1-5_split-einsum_cn models into the path below and set it as the url: StableDiffusionPipeline(resourcesAt: url)

macOS: file:///Users/kyd6/MochiDiffusion/models/stable-diffusion-1-5-split_einsum_cn/

iOS: // Custom model put under "Download" folder file:///private/var/mobile/Containers/Shared/AppGroup/C73F1304-A597-4223-9538-0EBF0AF8800F/File%20Provider%20Storage/Downloads/Stable-Diffusion-v1-5_split-einsum_cn/

// Default Hub Downloaded model put under Application folder file:///var/mobile/Containers/Data/Application/8C1DE5AA-76FC-4037-87DD-FCDE210BA66E/Library/Application%20Support/hf-diffusion-models/coreml-stable-diffusion-2-1-base-palettized_split_einsum_v2_compiled/

On macOS, it works well and has no problem; On iOS:

• it will crash at TextEncoder when generating images using Download folder
• It works find when using hub downloaded models in application folder.

The iOS throw Error below:

tableDiffusion/TextEncoder.swift:93: Fatal error: 'try!' expression unexpectedly raised an error: Error Domain=com.apple.CoreML Code=0 "Unable to load model: file:///private/var/mobile/Containers/Shared/AppGroup/C73F1304-A597-4223-9538-0EBF0AF8800F/File%20Provider%20Storage/Downloads/Stable-Diffusion-v1-5_split-einsum_cn/TextEncoder.mlmodelc/. Compile the model with Xcode or MLModel.compileModel(at:). " UserInfo={NSLocalizedDescription=Unable to load model: file:///private/var/mobile/Containers/Shared/AppGroup/C73F1304-A597-4223-9538-0EBF0AF8800F/File%20Provider%20Storage/Downloads/Stable-Diffusion-v1-5_split-einsum_cn/TextEncoder.mlmodelc/. Compile the model with Xcode or MLModel.compileModel(at:). , NSUnderlyingError=0x281a754d0 {Error Domain=com.apple.CoreML Code=3 "Failed to open file: /private/var/mobile/Containers/Shared/AppGroup/C73F1304-A597-4223-9538-0EBF0AF8800F/File Provider Storage/Downloads/Stable-Diffusion-v1-5_split-einsum_cn/TextEncoder.mlmodelc/coremldata.bin. It is not a valid .mlmodelc file. : unspecified iostream_category error" UserInfo={NSLocalizedDescription=Failed to open file: /private/var/mobile/Containers/Shared/AppGroup/C73F1304-A597-4223-9538-0EBF0AF8800F/File Provider Storage/Downloads/Stable-Diffusion-v1-5_split-einsum_cn/TextEncoder.mlmodelc/coremldata.bin. It is not a valid .mlmodelc file. : unspecified iostream_category error}}}

I don't know why it uses a folder located inside mlmodelc file in iOS, but when it comes to macOS it is OK.

How can I fix this issue?

[TimYao18]

I found if StableDiffusionPipeline(..., reduceMemory: true) will make this happen. Set reduceMemory to false, this will not happen.

But the app might be crashed due to out of memory.

[TimYao18]

In Image Generation with Swift: On iOS, the reduceMemory option should be set to false when constructing StableDiffusionPipeline

I think there might be something wrong?

[TimYao18]

Close issue due to not the same question about title.

[TimYao18]

Change the title and reopen this issue.