VisionConfig: add image normalization fields (mean, std, rescaleFactor)

Read per-channel normalization from metadata instead of hardcoding.
Fields are optional — bundles without them default to CLIP values
(the most common across VLMs). Gemma/SigLIP bundles specify their
own [0.5, 0.5, 0.5] values explicitly in metadata.json.

Author: stikves

swift/Sources/CoreAILanguageModels/Bundle/LanguageConfig.swift

@@ -160,22 +160,55 @@ public struct VisionConfig: Codable, Sendable, Equatable {
     /// Token ID used as a placeholder in the text sequence for image positions.
     public let imageTokenId: Int32
 
+    /// Per-channel normalization mean (RGB). Defaults to CLIP values when omitted.
+    public let imageMean: [Double]
+
+    /// Per-channel normalization std (RGB). Defaults to CLIP values when omitted.
+    public let imageStd: [Double]
+
+    /// Pixel rescale factor applied before normalization. Defaults to 1.0 when omitted.
+    public let rescaleFactor: Double
+
+    /// CLIP normalization (Qwen VL, Pixtral, InternVL, Phi-3.5-vision).
+    public static let clipMean = [0.48145466, 0.4578275, 0.40821073]
+    public static let clipStd = [0.26862954, 0.26130258, 0.27577711]
+
     public init(
         imageSize: Int,
         patchSize: Int,
         imageTokenCount: Int,
-        imageTokenId: Int32
+        imageTokenId: Int32,
+        imageMean: [Double]? = nil,
+        imageStd: [Double]? = nil,
+        rescaleFactor: Double? = nil
     ) {
         self.imageSize = imageSize
         self.patchSize = patchSize
         self.imageTokenCount = imageTokenCount
         self.imageTokenId = imageTokenId
+        self.imageMean = imageMean ?? Self.clipMean
+        self.imageStd = imageStd ?? Self.clipStd
+        self.rescaleFactor = rescaleFactor ?? 1.0
     }
 
     enum CodingKeys: String, CodingKey {
         case imageSize = "image_size"
         case patchSize = "patch_size"
         case imageTokenCount = "image_token_count"
         case imageTokenId = "image_token_id"
+        case imageMean = "image_mean"
+        case imageStd = "image_std"
+        case rescaleFactor = "rescale_factor"
+    }
+
+    public init(from decoder: Swift.Decoder) throws {
+        let c = try decoder.container(keyedBy: CodingKeys.self)
+        self.imageSize = try c.decode(Int.self, forKey: .imageSize)
+        self.patchSize = try c.decode(Int.self, forKey: .patchSize)
+        self.imageTokenCount = try c.decode(Int.self, forKey: .imageTokenCount)
+        self.imageTokenId = try c.decode(Int32.self, forKey: .imageTokenId)
+        self.imageMean = try c.decodeIfPresent([Double].self, forKey: .imageMean) ?? Self.clipMean
+        self.imageStd = try c.decodeIfPresent([Double].self, forKey: .imageStd) ?? Self.clipStd
+        self.rescaleFactor = try c.decodeIfPresent(Double.self, forKey: .rescaleFactor) ?? 1.0
     }
 }

swift/Sources/CoreAILanguageModels/InferenceEngines/CoreAISequentialVLMEngine.swift

@@ -312,16 +312,13 @@ public final class CoreAISequentialVLMEngine: MultimodalInferenceEngine, @unchec
         }
         self.llmFunction = llmFn
 
-        // Select image preprocessor based on vision config
-        // Default to gemma3/SigLIP preset; can be extended for other encoders
+        // Build image preprocessor from vision config normalization fields.
+        let vc = config.visionConfig
         self.imagePreprocessor = ImagePreprocessor(
-            targetSize: CGSize(
-                width: config.visionConfig.imageSize,
-                height: config.visionConfig.imageSize
-            ),
-            mean: (0.485, 0.456, 0.406),
-            std: (0.229, 0.224, 0.225),
-            rescaleFactor: 1.0
+            targetSize: CGSize(width: vc.imageSize, height: vc.imageSize),
+            mean: (CGFloat(vc.imageMean[0]), CGFloat(vc.imageMean[1]), CGFloat(vc.imageMean[2])),
+            std: (CGFloat(vc.imageStd[0]), CGFloat(vc.imageStd[1]), CGFloat(vc.imageStd[2])),
+            rescaleFactor: CGFloat(vc.rescaleFactor)
         )
 
         InstrumentsProfiler.endCustomInterval(