Add LaMa Core ML model for AI-powered inpainting

- Add LaMaFP16_512.mlpackage (~90MB) for high-quality object removal
- Add LaMaInpainter.swift wrapper with image preprocessing and merging
- Modify InpaintEngine to use LaMa first, gradient fill as fallback
- Fix brush mask size (use scale 1.0 instead of screen scale)
- Fix LaMa output size (use scale 1.0 in merge function)
- Add model loading wait with 5 second timeout

The LaMa model provides significantly better inpainting quality compared
to the gradient fill method, especially for complex backgrounds.

.gitignore

@@ -54,3 +54,4 @@ logs/
 # Temporary files
 *.tmp
 *.temp
+inpaint-ios-reference/

CheapRetouch/Features/Editor/BrushCanvasView.swift

@@ -205,8 +205,10 @@ struct BrushCanvasView: View {
             }
         }
 
-        // Create mask image from strokes
-        let renderer = UIGraphicsImageRenderer(size: imageSize)
+        // Create mask image from strokes (use scale 1.0 to match actual image pixels)
+        let format = UIGraphicsImageRendererFormat()
+        format.scale = 1.0  // Don't use screen scale, use actual pixel size
+        let renderer = UIGraphicsImageRenderer(size: imageSize, format: format)
         let maskImage = renderer.image { ctx in
             // Fill with black (not masked)
             UIColor.black.setFill()

CheapRetouch/Services/InpaintEngine/InpaintEngine.swift

@@ -74,14 +74,28 @@ actor InpaintEngine {
             throw InpaintError.memoryPressure
         }
 
+        // Try LaMa AI inpainting first for best quality
+        do {
+            DebugLogger.log("Attempting LaMa AI inpainting...")
+            let lamaResult = try await lamaInpainter.inpaint(image: image, mask: mask)
+            DebugLogger.imageInfo("Inpaint result (LaMa)", image: lamaResult)
+            return lamaResult
+        } catch {
+            DebugLogger.log("LaMa failed: \(error.localizedDescription), falling back to gradient fill")
+        }
+
+        // Fallback to gradient-based Metal inpainting
         if isMetalAvailable {
-            DebugLogger.log("Using Metal for inpainting")
+            DebugLogger.log("Using Metal gradient fill for inpainting")
             return try await inpaintWithMetal(image: image, mask: mask, isPreview: false)
         } else {
             DebugLogger.log("Using Accelerate fallback for inpainting")
             return try await inpaintWithAccelerate(image: image, mask: mask)
         }
     }
+    
+    /// LaMa inpainter for AI-powered inpainting (initialized lazily on first use)
+    private nonisolated(unsafe) lazy var lamaInpainter: LaMaInpainter = LaMaInpainter()
 
     func inpaintPreview(image: CGImage, mask: CGImage) async throws -> CGImage {
         // Scale down for preview if needed

CheapRetouch/Services/InpaintEngine/LaMaFP16_512.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "834F1D4D-C413-4927-9314-FF1187E2F6B4": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "DFC457AE-DC53-4BC5-B66A-1A6B1CB59064": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "DFC457AE-DC53-4BC5-B66A-1A6B1CB59064"
+}

CheapRetouch/Services/InpaintEngine/LaMaInpainter.swift

@@ -0,0 +1,346 @@
+//
+//  LaMaInpainter.swift
+//  CheapRetouch
+//
+//  LaMa (Large Mask Inpainting) Core ML wrapper for AI-powered object removal.
+//  Based on https://github.com/wudijimao/Inpaint-iOS
+//
+
+import UIKit
+import CoreML
+
+/// LaMa-based inpainting using Core ML
+/// Provides high-quality AI-powered inpainting for object removal
+final class LaMaInpainter {
+    
+    /// Fixed input size for the LaMa model
+    private let modelSize: Int = 512
+    
+    /// The Core ML model instance
+    private var model: LaMaFP16_512?
+    
+    /// Configuration for the model
+    private let config: MLModelConfiguration
+    
+    /// Work queue for inference
+    private let workQueue = DispatchQueue(label: "com.cheapretouch.lama", qos: .userInitiated)
+    
+    /// CI context for image operations
+    private let ciContext = CIContext()
+    
+    init() {
+        config = MLModelConfiguration()
+        config.computeUnits = .cpuAndGPU
+        preloadModel()
+    }
+    
+    /// Preload the model in the background
+    private func preloadModel() {
+        workQueue.async { [weak self] in
+            guard let self = self else { return }
+            do {
+                self.model = try LaMaFP16_512(configuration: self.config)
+                DebugLogger.log("LaMa model loaded successfully")
+            } catch {
+                DebugLogger.error("Failed to load LaMa model: \(error)")
+            }
+        }
+    }
+    
+    /// Check if the model is ready
+    var isModelReady: Bool {
+        return model != nil
+    }
+    
+    /// Inpaint the masked region of an image using LaMa
+    /// - Parameters:
+    ///   - image: The source CGImage
+    ///   - mask: The mask CGImage (white = areas to inpaint)
+    /// - Returns: The inpainted CGImage
+    func inpaint(image: CGImage, mask: CGImage) async throws -> CGImage {
+        // Wait for model to be ready (with timeout)
+        let maxWaitTime: TimeInterval = 5.0
+        let startTime = Date()
+        
+        while model == nil {
+            if -startTime.timeIntervalSinceNow > maxWaitTime {
+                throw LaMaError.modelNotLoaded
+            }
+            try await Task.sleep(nanoseconds: 100_000_000) // 100ms
+        }
+        
+        guard let model = model else {
+            throw LaMaError.modelNotLoaded
+        }
+        
+        let originalSize = CGSize(width: image.width, height: image.height)
+        
+        return try await withCheckedThrowingContinuation { continuation in
+            workQueue.async {
+                do {
+                    let result = try self.performInpainting(
+                        model: model,
+                        image: image,
+                        mask: mask,
+                        originalSize: originalSize
+                    )
+                    continuation.resume(returning: result)
+                } catch {
+                    continuation.resume(throwing: error)
+                }
+            }
+        }
+    }
+    
+    /// Perform the actual inpainting operation
+    private func performInpainting(
+        model: LaMaFP16_512,
+        image: CGImage,
+        mask: CGImage,
+        originalSize: CGSize
+    ) throws -> CGImage {
+        
+        DebugLogger.processing("LaMa inpainting started")
+        let startTime = Date()
+        
+        // Find the bounding box of the mask to crop efficiently
+        let maskBounds = findMaskBounds(mask: mask)
+        DebugLogger.log("Mask bounds: \(maskBounds)")
+        
+        // Calculate the crop region (expand to square and add padding)
+        let cropRegion = calculateCropRegion(
+            maskBounds: maskBounds,
+            imageSize: originalSize,
+            targetSize: modelSize
+        )
+        DebugLogger.log("Crop region: \(cropRegion)")
+        
+        // Crop the image and mask to the region
+        guard let croppedImage = cropImage(image, to: cropRegion),
+              let croppedMask = cropImage(mask, to: cropRegion) else {
+            throw LaMaError.imageProcessingFailed
+        }
+        
+        // Resize to model input size
+        guard let resizedImage = resizeImage(croppedImage, to: CGSize(width: modelSize, height: modelSize)),
+              let resizedMask = resizeImage(croppedMask, to: CGSize(width: modelSize, height: modelSize)) else {
+            throw LaMaError.imageProcessingFailed
+        }
+        
+        // Convert to pixel buffers
+        guard let imageBuffer = createPixelBuffer(from: resizedImage, format: kCVPixelFormatType_32ARGB),
+              let maskBuffer = createGrayscalePixelBuffer(from: resizedMask) else {
+            throw LaMaError.bufferCreationFailed
+        }
+        
+        // Run inference
+        DebugLogger.log("Running LaMa inference...")
+        let output = try model.prediction(image: imageBuffer, mask: maskBuffer)
+        
+        // Convert output to CGImage
+        guard let outputImage = cgImageFromPixelBuffer(output.output) else {
+            throw LaMaError.outputConversionFailed
+        }
+        
+        // Resize output back to crop region size
+        guard let resizedOutput = resizeImage(outputImage, to: cropRegion.size) else {
+            throw LaMaError.imageProcessingFailed
+        }
+        
+        // Merge back into original image
+        let finalImage = mergeIntoOriginal(
+            original: image,
+            inpainted: resizedOutput,
+            at: cropRegion.origin
+        )
+        
+        let elapsed = -startTime.timeIntervalSinceNow
+        DebugLogger.processing("LaMa inpainting completed in \(String(format: "%.2f", elapsed))s")
+        
+        return finalImage
+    }
+    
+    // MARK: - Helper Methods
+    
+    /// Find the bounding box of white pixels in the mask
+    private func findMaskBounds(mask: CGImage) -> CGRect {
+        let width = mask.width
+        let height = mask.height
+        
+        guard let data = mask.dataProvider?.data,
+              let bytes = CFDataGetBytePtr(data) else {
+            return CGRect(x: 0, y: 0, width: CGFloat(width), height: CGFloat(height))
+        }
+        
+        var minX = width, minY = height, maxX = 0, maxY = 0
+        let bytesPerPixel = mask.bitsPerPixel / 8
+        let bytesPerRow = mask.bytesPerRow
+        
+        for y in 0..<height {
+            for x in 0..<width {
+                let offset = y * bytesPerRow + x * bytesPerPixel
+                let value = bytes[offset]
+                if value > 127 {
+                    minX = min(minX, x)
+                    minY = min(minY, y)
+                    maxX = max(maxX, x)
+                    maxY = max(maxY, y)
+                }
+            }
+        }
+        
+        if minX > maxX || minY > maxY {
+            // No white pixels found, return full image
+            return CGRect(x: 0, y: 0, width: CGFloat(width), height: CGFloat(height))
+        }
+        
+        return CGRect(
+            x: CGFloat(minX),
+            y: CGFloat(minY),
+            width: CGFloat(maxX - minX + 1),
+            height: CGFloat(maxY - minY + 1)
+        )
+    }
+    
+    /// Calculate the crop region for the mask area
+    private func calculateCropRegion(maskBounds: CGRect, imageSize: CGSize, targetSize: Int) -> CGRect {
+        // Add 20% padding around the mask
+        let padding = max(maskBounds.width, maskBounds.height) * 0.2
+        var region = maskBounds.insetBy(dx: -padding, dy: -padding)
+        
+        // Make it square (use the larger dimension)
+        let maxSide = max(region.width, region.height)
+        let centerX = region.midX
+        let centerY = region.midY
+        region = CGRect(
+            x: centerX - maxSide / 2,
+            y: centerY - maxSide / 2,
+            width: maxSide,
+            height: maxSide
+        )
+        
+        // Ensure minimum size matches model size
+        if region.width < CGFloat(targetSize) {
+            let diff = CGFloat(targetSize) - region.width
+            region = region.insetBy(dx: -diff / 2, dy: -diff / 2)
+        }
+        
+        // Clamp to image bounds
+        region.origin.x = max(0, min(region.origin.x, imageSize.width - region.width))
+        region.origin.y = max(0, min(region.origin.y, imageSize.height - region.height))
+        
+        // Ensure we don't exceed image bounds
+        if region.maxX > imageSize.width {
+            region.origin.x = imageSize.width - region.width
+        }
+        if region.maxY > imageSize.height {
+            region.origin.y = imageSize.height - region.height
+        }
+        
+        // Final clamp if region is larger than image
+        region.origin.x = max(0, region.origin.x)
+        region.origin.y = max(0, region.origin.y)
+        region.size.width = min(region.width, imageSize.width)
+        region.size.height = min(region.height, imageSize.height)
+        
+        return CGRect(
+            x: floor(region.origin.x),
+            y: floor(region.origin.y),
+            width: ceil(region.width),
+            height: ceil(region.height)
+        )
+    }
+    
+    /// Crop an image to the specified region
+    private func cropImage(_ image: CGImage, to rect: CGRect) -> CGImage? {
+        return image.cropping(to: rect)
+    }
+    
+    /// Resize an image to the specified size
+    private func resizeImage(_ image: CGImage, to size: CGSize) -> CGImage? {
+        let ciImage = CIImage(cgImage: image)
+        let scaleX = size.width / CGFloat(image.width)
+        let scaleY = size.height / CGFloat(image.height)
+        let scaled = ciImage.transformed(by: CGAffineTransform(scaleX: scaleX, y: scaleY))
+        return ciContext.createCGImage(scaled, from: scaled.extent)
+    }
+    
+    /// Create an ARGB pixel buffer from a CGImage
+    private func createPixelBuffer(from image: CGImage, format: OSType) -> CVPixelBuffer? {
+        do {
+            let feature = try MLFeatureValue(
+                cgImage: image,
+                pixelsWide: modelSize,
+                pixelsHigh: modelSize,
+                pixelFormatType: format
+            )
+            return feature.imageBufferValue
+        } catch {
+            DebugLogger.error("Failed to create pixel buffer: \(error)")
+            return nil
+        }
+    }
+    
+    /// Create a grayscale pixel buffer from a CGImage
+    private func createGrayscalePixelBuffer(from image: CGImage) -> CVPixelBuffer? {
+        do {
+            let feature = try MLFeatureValue(
+                cgImage: image,
+                pixelsWide: modelSize,
+                pixelsHigh: modelSize,
+                pixelFormatType: kCVPixelFormatType_OneComponent8
+            )
+            return feature.imageBufferValue
+        } catch {
+            DebugLogger.error("Failed to create grayscale buffer: \(error)")
+            return nil
+        }
+    }
+    
+    /// Convert a CVPixelBuffer to CGImage
+    private func cgImageFromPixelBuffer(_ buffer: CVPixelBuffer) -> CGImage? {
+        let ciImage = CIImage(cvPixelBuffer: buffer)
+        return ciContext.createCGImage(ciImage, from: ciImage.extent)
+    }
+    
+    /// Merge the inpainted region back into the original image
+    private func mergeIntoOriginal(original: CGImage, inpainted: CGImage, at position: CGPoint) -> CGImage {
+        let size = CGSize(width: original.width, height: original.height)
+        let inpaintedSize = CGSize(width: inpainted.width, height: inpainted.height)
+        
+        // Use scale 1.0 to match actual pixel size (not screen scale)
+        let format = UIGraphicsImageRendererFormat()
+        format.scale = 1.0
+        let renderer = UIGraphicsImageRenderer(size: size, format: format)
+        let resultImage = renderer.image { context in
+            // Draw original
+            UIImage(cgImage: original).draw(at: .zero)
+            // Draw inpainted region on top
+            UIImage(cgImage: inpainted).draw(in: CGRect(origin: position, size: inpaintedSize))
+        }
+        
+        return resultImage.cgImage!
+    }
+}
+
+// MARK: - Errors
+
+enum LaMaError: Error, LocalizedError {
+    case modelNotLoaded
+    case imageProcessingFailed
+    case bufferCreationFailed
+    case outputConversionFailed
+    
+    var errorDescription: String? {
+        switch self {
+        case .modelNotLoaded:
+            return "LaMa model is not loaded"
+        case .imageProcessingFailed:
+            return "Failed to process image for inpainting"
+        case .bufferCreationFailed:
+            return "Failed to create pixel buffer"
+        case .outputConversionFailed:
+            return "Failed to convert model output to image"
+        }
+    }
+}