v0.3.0 - Python Pipelines

Python Pipelines

You can now run Diffusers Python pipelines directly from C#, with everything managed for you.

• Automatic Python Setup
  Downloads and installs an isolated Python runtime — no system Python required.

• Device-Specific Virtual Environments
  Creates a dedicated virtual environment tailored to the selected device.

• Hugging Face Model Downloads
  Automatically downloads Diffusers models from Hugging Face.

• LoRA Support
  Load and apply LoRA models in supported pipelines.

• CUDA Support
  Support for CUDA devices with support for ROCM in development

• Window Demo App
  Basic Window PythonDemo App available for download below

Supported Diffusers Pipelines

Chroma

• ChromaPipeline
• ChromaImg2ImgPipeline

Qwen Image

• QwenImagePipeline
• QwenImageImg2ImgPipeline
• QwenImageEditPipeline

Wan Video

• WanPipeline
• WanImageToVideoPipeline

Z-Image

• ZImagePipeline
• ZImageImg2ImgPipeline

---

Basic Example

Python Virtual Environment

The PythonManager can be used to download and install Python and create virtual environments

// Virtual Environment Config
var serverConfig = new EnvironmentConfig
{
    Environment = "default-cuda",
    Directory = "PythonRuntime",
    Requirements =
    [
        "--extra-index-url https://download.pytorch.org/whl/cu118",
        "torch==2.7.0+cu118",
        "typing",
        "wheel",
        "transformers",
        "accelerate",
        "diffusers",
        "protobuf",
        "sentencepiece",
        "pillow",
        "ftfy",
        "scipy",
        "peft"
    ]
};

// PythonManager
var pythonManager = new PythonManager(serverConfig);

// Create/Load Virtual Environment
await pythonManager .CreateEnvironmentAsync(PipelineProgress.ConsoleCallback);

Python Pipelines

Once you have created a virtual environment you can now load a pipeline

// Pipeline Config
var pipelineConfig = new PipelineConfig
{
    Path = "Qwen/Qwen-Image-Edit",
    Pipeline = "QwenImagePipeline",
    ProcessType = ProcessType.ImageEdit,
    IsFullOffloadEnabled = true,
    DataType = DataType.Bfloat16
};

// Create Pipeline Proxy
using (var pythonPipeline = new PythonPipeline(pipelineConfig, PipelineProgress.ConsoleCallback))
{
    // Download/Load Model
    await pythonPipeline.LoadAsync();

    // Generate Option
    var options = new PipelineOptions
    {
        Prompt = "Yarn art style",
        Steps = 30,
        Width = 1024,
        Height = 1024,
        GuidanceScale = 4f,
        Scheduler = SchedulerType.FlowMatchEulerDiscrete,
        ImageInput = new ImageInput("Image.png")
    };

    // Generate
    var response = await pythonPipeline.GenerateAsync(options);

    // Save Image
    await response
        .AsImageTensor()
        .SaveAsync("Result.png");
}

Full Changelog: v0.2.2...v0.3.0