LMSupply.ImageGenerator 0.10.0

LMSupply.ImageGenerator

A simple .NET library for local text-to-image generation using Latent Consistency Models (LCM). Designed for edge deployment with fast 2-4 step inference.

Features

• Fast Generation: LCM enables 2-4 step inference (vs 20-50 for standard diffusion)
• Edge-Optimized: ~3-4GB VRAM requirement, suitable for consumer GPUs
• GPU Acceleration: CUDA, DirectML, and CoreML support via ONNX Runtime
• Lazy Downloads: Models downloaded on-demand from HuggingFace
• Streaming Progress: Real-time step-by-step progress with optional previews

Quick Start

using LMSupply.ImageGenerator;

// Load the default LCM model
await using var generator = await LocalImageGenerator.LoadAsync("default");

// Generate an image
var result = await generator.GenerateAsync("A sunset over mountains");
await result.SaveAsync("output.png");

Model Aliases

Generation Options

var options = new GenerationOptions
{
    Width = 512,           // Image width (must be divisible by 8)
    Height = 512,          // Image height (must be divisible by 8)
    Steps = 4,             // Inference steps (2-4 for LCM)
    GuidanceScale = 1.0f,  // CFG scale (1.0-2.0 for LCM)
    Seed = 42,             // Random seed for reproducibility
    NegativePrompt = "blurry, low quality",
    GeneratePreviews = true // Enable step-by-step previews
};

var result = await generator.GenerateAsync("A cat wearing a hat", options);

Streaming Generation

await foreach (var step in generator.GenerateStreamingAsync("A beautiful landscape"))
{
    Console.WriteLine($"Step {step.StepNumber}/{step.TotalSteps} ({step.Progress:F1}%)");

    if (step.HasPreview)
    {
        // Save intermediate preview
        await File.WriteAllBytesAsync($"preview_{step.StepNumber}.png", step.PreviewData!);
    }

    if (step.IsFinal && step.FinalImage != null)
    {
        await step.FinalImage.SaveAsync("final.png");
    }
}

Batch Generation

// Generate 4 variations with different seeds
var images = await generator.GenerateBatchAsync("A fantasy castle", count: 4);

for (int i = 0; i < images.Length; i++)
{
    await images[i].SaveAsync($"castle_{i}.png");
    Console.WriteLine($"Image {i}: Seed={images[i].Seed}, Time={images[i].GenerationTime.TotalSeconds:F2}s");
}

GPU Configuration

var options = new ImageGeneratorOptions
{
    Provider = ExecutionProvider.Cuda,  // Force CUDA (auto-detected by default)
    DeviceId = 0,                       // GPU device index
    UseFp16 = true,                     // Use FP16 for reduced memory
    CacheDirectory = "/path/to/cache"   // Custom model cache location
};

await using var generator = await LocalImageGenerator.LoadAsync("default", options);

Requirements

• .NET 10.0+
• GPU with 3-4GB VRAM (recommended)
• Supported execution providers:
  • CUDA 11/12 (NVIDIA GPUs)
  • DirectML (Windows, any DirectX 12 GPU)
  • CoreML (macOS, Apple Silicon)
  • CPU (fallback, slower)

How It Works

LMSupply.ImageGenerator uses Latent Consistency Models (LCM), a distillation technique that enables high-quality image generation in just 2-4 steps instead of the traditional 20-50 steps required by Stable Diffusion.

The pipeline consists of:

1. CLIP Text Encoder: Converts text prompts to embeddings
2. LCM UNet: Denoises latents in 2-4 steps
3. VAE Decoder: Converts latents to pixel space
4. LCM Scheduler: Manages the accelerated diffusion process

License

MIT License - See LICENSE file for details.