diff --git a/README.md b/README.md
index 2e58656be..e86dfac80 100644
--- a/README.md
+++ b/README.md
@@ -155,13 +155,17 @@ mask = mask_utils.decode(annotation["segmentation"])
 
 See [here](https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocotools/mask.py) for more instructions to manipulate masks stored in RLE format.
 
+## Note on Reproducibility with GPU
+
+The fine-tuning training process is not reproducible when using GPU. Using `torch.use_deterministic_algorithms(True)` with GPU does not work due to non-deterministic routines in PyTorch/CUDA. The `ResizeLongestSide` class in `segment_anything/utils/transforms.py` is suspected to be the cause of non-reproducibility. The current setup with `torch.use_deterministic_algorithms(False)` does not ensure reproducibility in fine-tuning training with GPU. As a potential workaround, you can use CPU for reproducibility, although it may be significantly slower.
+
 ## License
 
 The model is licensed under the [Apache 2.0 license](LICENSE).
 
 ## Contributing
 
-See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF_CONDUCT.md).
+See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF-CONDUCT.md).
 
 ## Contributors
 
diff --git a/notebooks/onnx_model_example.ipynb b/notebooks/onnx_model_example.ipynb
index ca49c3571..6b5f5daca 100644
--- a/notebooks/onnx_model_example.ipynb
+++ b/notebooks/onnx_model_example.ipynb
@@ -535,6 +535,7 @@
    "id": "d778a8fb",
    "metadata": {},
    "outputs": [],
+   "source": [],
    "source": [
     "masks.shape"
    ]
@@ -748,6 +749,16 @@
     "plt.axis('off')\n",
     "plt.show()"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d3e8b1b2",
+   "metadata": {},
+   "source": [
+    "## Note on Reproducibility with GPU\n",
+    "\n",
+    "The fine-tuning training process is not reproducible when using GPU. Using `torch.use_deterministic_algorithms(True)` with GPU does not work due to non-deterministic routines in PyTorch/CUDA. The `ResizeLongestSide` class in `segment_anything/utils/transforms.py` is suspected to be the cause of non-reproducibility. The current setup with `torch.use_deterministic_algorithms(False)` does not ensure reproducibility in fine-tuning training with GPU. As a potential workaround, you can use CPU for reproducibility, although it may be significantly slower."
+   ]
   }
  ],
  "metadata": {
@@ -772,3 +783,5 @@
  "nbformat": 4,
  "nbformat_minor": 5
 }
+    "    ax.scatter(neg_points[:, 0], neg_points[:, 1], color='red', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)   \n",
+    "    \n",
diff --git a/scripts/amg.py b/scripts/amg.py
index f2dbf676a..ba35736a1 100644
--- a/scripts/amg.py
+++ b/scripts/amg.py
@@ -218,6 +218,9 @@ def main(args: argparse.Namespace) -> None:
             continue
         image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
+        # Use apply_image_deterministic for deterministic resizing
+        image = generator.predictor.transform.apply_image_deterministic(image)
+
         masks = generator.generate(image)
 
         base = os.path.basename(t)
diff --git a/segment_anything/utils/transforms.py b/segment_anything/utils/transforms.py
index c08ba1e3d..fc54b133a 100644
--- a/segment_anything/utils/transforms.py
+++ b/segment_anything/utils/transforms.py
@@ -30,6 +30,17 @@ def apply_image(self, image: np.ndarray) -> np.ndarray:
         target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
         return np.array(resize(to_pil_image(image), target_size))
 
+    def apply_image_deterministic(self, image: np.ndarray) -> np.ndarray:
+        """
+        Expects a numpy array with shape HxWxC in uint8 format.
+        Uses torch.nn.functional.interpolate with mode='nearest' for deterministic resizing.
+        """
+        target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
+        image_torch = torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float()
+        resized_image_torch = F.interpolate(image_torch, size=target_size, mode='nearest')
+        resized_image = resized_image_torch.squeeze(0).permute(1, 2, 0).byte().numpy()
+        return resized_image
+
     def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
         """
         Expects a numpy array of length 2 in the final dimension. Requires the