diff --git a/tests/Furnace.Tests/Furnace.Tests.fsproj b/tests/Furnace.Tests/Furnace.Tests.fsproj
index d5b940a7..8f647df2 100644
--- a/tests/Furnace.Tests/Furnace.Tests.fsproj
+++ b/tests/Furnace.Tests/Furnace.Tests.fsproj
@@ -33,6 +33,7 @@
     <Compile Include="TestRandom.fs" />
     <Compile Include="TestDistributions.fs" />
     <Compile Include="TestExtensions.fs" />
+    <Compile Include="TestTorchExtensions.fs" />
     <Compile Include="TestData.fs" />
     <Compile Include="TestModel.fs" />
     <Compile Include="TestModel.BatchNorm.fs" />
diff --git a/tests/Furnace.Tests/TestTorchExtensions.fs b/tests/Furnace.Tests/TestTorchExtensions.fs
new file mode 100644
index 00000000..d36b992d
--- /dev/null
+++ b/tests/Furnace.Tests/TestTorchExtensions.fs
@@ -0,0 +1,188 @@
+// Copyright (c) 2016-     University of Oxford (Atılım Güneş Baydin <gunes@robots.ox.ac.uk>)
+// and other contributors, see LICENSE in root of repository.
+//
+// BSD 2-Clause License. See LICENSE in root of repository.
+
+namespace Tests
+
+open NUnit.Framework
+open Furnace
+open Furnace.Backends.Torch
+
+[<TestFixture>]
+module TestTorchExtensions =
+
+    [<Test>]
+    let TestTorchExtensionsFromTorchTensor() =
+        use config = FurnaceImage.useConfig(backend=Backend.Torch, dtype=Dtype.Float32)
+        
+        // Create a simple tensor with known values
+        let values = [| 1.0f; 2.0f; 3.0f; 4.0f |]
+        let originalTensor = FurnaceImage.tensor(values)
+        
+        // Convert to torch tensor and back
+        let torchTensor = originalTensor.toTorch()
+        let reconstructedTensor = FurnaceImage.fromTorch(torchTensor)
+        
+        // Verify the reconstructed tensor has the same values
+        Assert.CheckEqual(originalTensor.shape, reconstructedTensor.shape)
+        Assert.CheckEqual(originalTensor.dtype, reconstructedTensor.dtype)
+        Assert.CheckEqual(originalTensor.backend, reconstructedTensor.backend)
+        
+        // Check values are preserved
+        let originalArray = originalTensor.toArray() :?> float32[]
+        let reconstructedArray = reconstructedTensor.toArray() :?> float32[]
+        
+        Assert.CheckEqual(originalArray.Length, reconstructedArray.Length)
+        for i in 0 .. originalArray.Length - 1 do
+            Assert.AreEqual(float originalArray[i], float reconstructedArray[i], 1e-6)
+
+    [<Test>]
+    let TestTorchExtensionsFromTorchTensor2D() =
+        use config = FurnaceImage.useConfig(backend=Backend.Torch, dtype=Dtype.Float32)
+        
+        // Create a 2D tensor
+        let values = [| [| 1.0f; 2.0f |]; [| 3.0f; 4.0f |] |]
+        let originalTensor = FurnaceImage.tensor(values)
+        
+        // Convert to torch tensor and back
+        let torchTensor = originalTensor.toTorch()
+        let reconstructedTensor = FurnaceImage.fromTorch(torchTensor)
+        
+        // Verify shape and properties
+        Assert.CheckEqual([| 2; 2 |], reconstructedTensor.shape)
+        Assert.CheckEqual(Dtype.Float32, reconstructedTensor.dtype)
+        Assert.CheckEqual(Backend.Torch, reconstructedTensor.backend)
+        
+        // Check values (2D tensor returns float32[,])
+        let originalArray = originalTensor.toArray() :?> float32[,]
+        let reconstructedArray = reconstructedTensor.toArray() :?> float32[,]
+        
+        Assert.CheckEqual(originalArray.GetLength(0), reconstructedArray.GetLength(0))
+        Assert.CheckEqual(originalArray.GetLength(1), reconstructedArray.GetLength(1))
+        Assert.AreEqual(1.0, float reconstructedArray[0,0], 1e-6)
+        Assert.AreEqual(2.0, float reconstructedArray[0,1], 1e-6)
+        Assert.AreEqual(3.0, float reconstructedArray[1,0], 1e-6)
+        Assert.AreEqual(4.0, float reconstructedArray[1,1], 1e-6)
+
+    [<Test>]
+    let TestTorchExtensionsFromTorchScalar() =
+        use config = FurnaceImage.useConfig(backend=Backend.Torch, dtype=Dtype.Float32)
+        
+        // Create a scalar tensor
+        let originalScalar = FurnaceImage.scalar(42.0f)
+        
+        // Convert to torch tensor and back
+        let torchTensor = originalScalar.toTorch()
+        let reconstructedScalar = FurnaceImage.fromTorch(torchTensor)
+        
+        // Verify scalar properties
+        Assert.CheckEqual([||], reconstructedScalar.shape)  // Scalar has empty shape
+        Assert.CheckEqual(Dtype.Float32, reconstructedScalar.dtype)
+        Assert.CheckEqual(Backend.Torch, reconstructedScalar.backend)
+        
+        // Check the scalar value
+        let originalValue = originalScalar.toScalar() :?> float32
+        let reconstructedValue = reconstructedScalar.toScalar() :?> float32
+        Assert.AreEqual(42.0, float reconstructedValue, 1e-6)
+        Assert.AreEqual(float originalValue, float reconstructedValue, 1e-6)
+
+    [<Test>]
+    let TestTorchExtensionsToTorchWithDifferentDtypes() =
+        use config = FurnaceImage.useConfig(backend=Backend.Torch)
+        
+        // Test with different data types
+        let floatTensor = FurnaceImage.tensor([| 1.5f; 2.5f |], dtype=Dtype.Float32)
+        let doubleTensor = FurnaceImage.tensor([| 1.5; 2.5 |], dtype=Dtype.Float64)
+        let intTensor = FurnaceImage.tensor([| 1; 2 |], dtype=Dtype.Int32)
+        
+        // Convert to torch tensors
+        let torchFloat = floatTensor.toTorch()
+        let torchDouble = doubleTensor.toTorch()
+        let torchInt = intTensor.toTorch()
+        
+        // Verify the torch tensors exist and have correct properties
+        Assert.IsNotNull(torchFloat)
+        Assert.IsNotNull(torchDouble)
+        Assert.IsNotNull(torchInt)
+        
+        // Convert back and verify
+        let reconstructedFloat = FurnaceImage.fromTorch(torchFloat)
+        let reconstructedDouble = FurnaceImage.fromTorch(torchDouble)
+        let reconstructedInt = FurnaceImage.fromTorch(torchInt)
+        
+        Assert.CheckEqual(Dtype.Float32, reconstructedFloat.dtype)
+        Assert.CheckEqual(Dtype.Float64, reconstructedDouble.dtype)
+        Assert.CheckEqual(Dtype.Int32, reconstructedInt.dtype)
+
+    [<Test>]
+    let TestTorchExtensionsToTorchWithNonTorchBackend() =
+        use config = FurnaceImage.useConfig(backend=Backend.Reference, dtype=Dtype.Float32)
+        
+        // Create a tensor with Reference backend
+        let referenceTensor = FurnaceImage.tensor([| 1.0f; 2.0f; 3.0f |])
+        
+        // Verify it's using Reference backend
+        Assert.CheckEqual(Backend.Reference, referenceTensor.backend)
+        
+        // Attempting to convert to torch should throw an exception
+        Assert.Throws<System.Exception>(fun () -> 
+            let torchTensor = referenceTensor.toTorch()
+            torchTensor |> ignore
+        ) |> ignore
+
+    [<Test>]
+    let TestTorchExtensionsRoundTripLargerTensor() =
+        use config = FurnaceImage.useConfig(backend=Backend.Torch, dtype=Dtype.Float32)
+        
+        // Create a larger 3D tensor
+        let originalTensor = FurnaceImage.randn([2; 3; 4])
+        
+        // Convert to torch and back
+        let torchTensor = originalTensor.toTorch()
+        let reconstructedTensor = FurnaceImage.fromTorch(torchTensor)
+        
+        // Verify properties match
+        Assert.CheckEqual(originalTensor.shape, reconstructedTensor.shape)
+        Assert.CheckEqual(originalTensor.dtype, reconstructedTensor.dtype)
+        Assert.CheckEqual(originalTensor.backend, reconstructedTensor.backend)
+        Assert.CheckEqual(24, reconstructedTensor.nelement)  // 2 * 3 * 4 = 24
+        
+        // Verify values are preserved (3D tensor returns float32[,,])
+        let originalArray = originalTensor.toArray() :?> float32[,,]
+        let reconstructedArray = reconstructedTensor.toArray() :?> float32[,,]
+        
+        Assert.CheckEqual(originalArray.GetLength(0), reconstructedArray.GetLength(0))
+        Assert.CheckEqual(originalArray.GetLength(1), reconstructedArray.GetLength(1))
+        Assert.CheckEqual(originalArray.GetLength(2), reconstructedArray.GetLength(2))
+        
+        // Check values are preserved (allowing for floating point precision)
+        for i in 0 .. originalArray.GetLength(0) - 1 do
+            for j in 0 .. originalArray.GetLength(1) - 1 do
+                for k in 0 .. originalArray.GetLength(2) - 1 do
+                    Assert.AreEqual(float originalArray[i,j,k], float reconstructedArray[i,j,k], 1e-6)
+
+    [<Test>]
+    let TestTorchExtensionsBoolTensorRoundTrip() =
+        use config = FurnaceImage.useConfig(backend=Backend.Torch, dtype=Dtype.Float32)
+        
+        // Create a boolean tensor (need to specify dtype explicitly)
+        let originalTensor = FurnaceImage.tensor([| true; false; true; false |], dtype=Dtype.Bool)
+        
+        // Convert to torch and back
+        let torchTensor = originalTensor.toTorch()
+        let reconstructedTensor = FurnaceImage.fromTorch(torchTensor)
+        
+        // Verify properties
+        Assert.CheckEqual(originalTensor.shape, reconstructedTensor.shape)
+        Assert.CheckEqual(Dtype.Bool, reconstructedTensor.dtype)
+        
+        // Check boolean values are preserved
+        let originalArray = originalTensor.toArray() :?> bool[]
+        let reconstructedArray = reconstructedTensor.toArray() :?> bool[]
+        
+        Assert.CheckEqual(4, reconstructedArray.Length)
+        Assert.CheckEqual(true, reconstructedArray[0])
+        Assert.CheckEqual(false, reconstructedArray[1])
+        Assert.CheckEqual(true, reconstructedArray[2])
+        Assert.CheckEqual(false, reconstructedArray[3])
\ No newline at end of file