模型类别

首先要下载模型,Stable Diffusion模型可以在huggingface或者Civitai下载到。但是在这两个网站上下载的模型可能会有三种格式。

CoreML格式

这种类别的模型较少,文件主要以.mlmodelc.mlmodel为主,其文件结构大致为:

1
2
3
4
5
6
├── TextEncoder.mlmodelc
├── TextEncoder2.mlmodelc
├── Unet.mlmodelc
├── VAEDecoder.mlmodelc
├── merges.txt
└── vocab.json

Diffusers格式

huggingface上下载的模型大多是这种类型,其文件结构大致为:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
├── model_index.json
├── scheduler
│   └── scheduler_config.json
├── text_encoder
│   ├── config.json
│   └── pytorch_model.bin
├── tokenizer
│   ├── merges.txt
│   ├── special_tokens_map.json
│   ├── tokenizer_config.json
│   └── vocab.json
├── unet
│   ├── config.json
│   └── diffusion_pytorch_model.bin
└── vae
    ├── config.json
    └── diffusion_pytorch_model.bin

safetensors格式

Civitai网站下载的大多是这种格式,就一个文件,非常方便。

模型转换

接下来,需要把下载下来的模型都转成CoreML格式,如果你在第一步下载的模型已经是CoreML格式,那么这一步就可以跳过。

Diffusers格式转CoreML格式

首先下载该仓库代码:ml-stable-diffusion。查看System Requirements检查自己的设备是否支持。然后安装依赖:

1
pip install -r requirements.txt

找到torch2coreml.py文件,执行以下命令:

1
2
3
4
5
6
7
8
9
python torch2coreml.py \
--bundle-resources-for-swift-cli \
--xl-version \
--convert-unet \
--convert-text-encoder \
--convert-vae-decoder \
--attention-implementation ORIGINAL \
--model-version /your/model/path \
-o /your/model/output/path

注意有个参数--xl-version,如果模型是sdxl类型的,就加上,否则把这行删除。另外如果你的模型支持图生图,你可以加上--convert-vae-encoder参数。

运行完该命令,应该在你指定的目录生成了文件,在Resources目录下的文件就是转换好的CoreML格式。

safetensors格式转Diffusers格式

首先下载该仓库代码:Diffusers。然后安装依赖:

1
pip install --upgrade diffusers

找到convert_original_stable_diffusion_to_diffusers.py文件并执行以下命令:

1
2
3
4
5
6
python convert_original_stable_diffusion_to_diffusers.py \
--checkpoint_path /your/model/path \
--dump_path /your/model/output/path \
--from_safetensors \
--half \
--device mps

这里--half表示转换时精度为fp16--device mps表示模型使用mps(GPU)进行推理。

运行完该命令,会生成Diffusers格式的模型,再利用Diffusers格式转CoreML格式的步骤,将模型转换为CoreML格式。

Swift调用Stable Diffusion模型

使用Huggingface提供的swift-coreml-diffusers库。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
import StableDiffusion
import CoreML

// 初始化CoreML配置
let config = MLModelConfiguration()
// 运行在GPU上(MAC限定)
config.computeUnits = MLComputeUnits.cpuAndGPU
// 初始化pipeline
var pipeline = try StableDiffusionPipeline(resourcesAt: modelDirectory,
                                      controlNet: [],
                                      configuration: config,
                                      reduceMemory: diffusersConfig.reduceMemory)
let pipeline.loadResources()
// 初始化图片推理配置
var pipelineConfig = StableDiffusionPipeline.Configuration(prompt: prompt)
pipelineConfig.stepCount = stepCount
pipelineConfig.guidanceScale = cfgScale
pipelineConfig.schedulerType = scheduler
// 开始图片推理
let images = try pipeline.generateImages(configuration: pipelineConfig,
                                          progressHandler: { progress in
})

我的AquariusAI项目提供了示例代码。

最后

那到底什么是CoreML呢?

Core ML 是Apple Silicon芯片产品(包括macOS、iOS、watchOS 和 tvOS)中使用的机器学习框架,用于执行快速预测或推理,在边缘轻松集成预训练的机器学习模型,从而可以对设备上的实时图像或视频进行实时预测。

Core ML 通过利用 CPU、GPU 和 神经网络引擎 ,同时最大程度地减小内存占用空间和功耗,来优化设备端性能。 由于模型严格地在用户设备上,因此无需任何网络连接,这有助于保护用户数据的私密性和 App 的响应速度。

简而言之,如果你的模型运行在Silicon芯片的苹果设备上,利用Core ML可以获得更快的性能和更低的内存及能耗。