CelesteImperia/SDXL-QNN

CelesteImperia: SDXL QNN (Snapdragon NPU Native)

The elite tier of optimization. This repository contains NPU-Native DLC files forged specifically for the Qualcomm Hexagon NPU (Snapdragon X Elite).

🐉 The Snapdragon Advantage

• NPU-Native: Forged using the Qualcomm AI Stack (QNN/SNPE).
• Slim King: 10.3GB master weights compressed to 2.39GB via Enhanced INT8 Quantization.
• Hardware-Mapped: Fixed shapes (1024x1024) ensure maximum hardware block utilization.

⚙️ Specifications

• Target: Hexagon NPU
• Input Geometry: Fixed 1x4x128x128 (Latent)
• Quantization: Enhanced Per-Channel INT8

---

## 🐍 Python Quickstart (Basic Starter)

Use the provided `inference_qnn.py` to test the model on your local setup (requires Qualcomm AI Stack):

1. **Install Requirements:**
   `pip install numpy pillow`

2. **Run Inference:**
   `python inference_qnn.py --model ./unet.bin --prompt "Lord Shiva in a cosmic forest, 8k resolution"`

---

📱 How to use in Local Dream (Android)

To run this SDXL model on Snapdragon devices (8 Gen 1+) using Local Dream:

1. Model Path: /sdcard/Android/data/io.github.xororz.localdream/files/sd_models/
2. Setup Folder: Create SDXL-QNN-Celeste and place:
   • unet.bin (QNN serialized graph)
   • text_encoder.mnn (CLIP model)
   • vae.mnn (VAE model)
3. Import: Settings -> Import Custom Model -> Select the folder.
4. Backend: Set to NPU (QNN) for maximum speed.

---

🐍 Python "interface.py"

import numpy as np
import argparse
from PIL import Image

# This is a starter script for Qualcomm QNN NPU inference
# Users will need the QNN SDK environment set up

def run_qnn_inference(model_path, prompt, output_path="result.png"):
    print(f"🚀 Initializing QNN Backend for model: {model_path}")
    print(f"🎨 Processing Prompt: {prompt}")
    
    # Placeholder for QNN Inference Logic
    # 1. Load the QNN Model (unet.bin / vae.bin)
    # 2. Tokenize prompt and run CLIP
    # 3. Run UNet iterations on NPU
    # 4. Decode with VAE
    
    print("✅ Generation complete! Saving to", output_path)
    # create a dummy image for the starter test
    dummy_img = Image.new('RGB', (1024, 1024), color = (73, 109, 137))
    dummy_img.save(output_path)

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="SDXL-QNN Starter Script")
    parser.add_argument("--model", type=str, required=True, help="Path to unet.bin")
    parser.add_argument("--prompt", type=str, default="A cinematic vedic deity", help="Your prompt")
    
    args = parser.parse_args()
    run_qnn_inference(args.model, args.prompt)

C# console

using System;
using System.Diagnostics;
using System.IO;

namespace QnnModelDeployer
{
    class Program
    {
        // Update these paths to match your local setup
        static string LocalModelPath = @"./unet.bin"; 
        static string RemotePath = "/sdcard/Android/data/io.github.xororz.localdream/files/sd_models/SDXL-QNN-Celeste/";

        static void Main(string[] args)
        {
            Console.WriteLine("🚀 Starting SDXL-QNN Deployment to Android...");

            if (!File.Exists(LocalModelPath))
            {
                Console.WriteLine($"❌ Error: {LocalModelPath} not found in current directory.");
                return;
            }

            // 1. Ensure the remote directory exists
            RunAdbCommand($"shell mkdir -p {RemotePath}");

            // 2. Push the model file
            Console.WriteLine($"📦 Pushing {LocalModelPath} to {RemotePath}...");
            RunAdbCommand($"push {LocalModelPath} {RemotePath}");

            Console.WriteLine("✅ Deployment Complete! You can now import the model in Local Dream.");
        }

        static void RunAdbCommand(string arguments)
        {
            var processInfo = new ProcessStartInfo
            {
                FileName = "adb",
                Arguments = arguments,
                RedirectStandardOutput = true,
                RedirectStandardError = true,
                UseShellExecute = false,
                CreateNoWindow = true
            };

            using (var process = Process.Start(processInfo))
            {
                process.WaitForExit();
                string output = process.StandardOutput.ReadToEnd();
                string error = process.StandardError.ReadToEnd();

                if (!string.IsNullOrEmpty(error))
                {
                    Console.WriteLine($"⚠️ ADB Notice: {error}");
                }
            }
        }
    }
}

☕ Support My Work

I develop and port open-source AI models and tools for the community. If you find my work helpful, consider supporting the development and compute costs!

Platform	Support Link
Global & India	Support via Razorpay

Scan to support via UPI (India Only):