diff --git a/cellpose/core.py b/cellpose/core.py
index ecfde40c..b3bcb092 100644
--- a/cellpose/core.py
+++ b/cellpose/core.py
@@ -52,7 +52,14 @@ def _use_gpu_torch(gpu_number=0):
     except:
         pass
     try:
-        device = torch.device('mps:' + str(gpu_number))
+        # ``gpu_number`` may be the string ``"mps"`` (passed straight through
+        # from ``--gpu_device mps``) rather than a device index, in which case
+        # ``"mps:mps"`` is not a valid device string. Fall back to the plain
+        # ``"mps"`` device for any non-integer gpu_number.
+        if str(gpu_number).isdigit():
+            device = torch.device('mps:' + str(gpu_number))
+        else:
+            device = torch.device('mps')
         _ = torch.zeros((1,1)).to(device)
         core_logger.info('** TORCH MPS version installed and working. **')
         return True
diff --git a/tests/test_import.py b/tests/test_import.py
index 2f4c709f..0e250cd6 100644
--- a/tests/test_import.py
+++ b/tests/test_import.py
@@ -21,6 +21,23 @@ def test_gpu_check():
     core.use_gpu()
 
 
+def test_assign_device_mps_string():
+    # Regression test for #1455: passing ``--gpu_device mps`` (i.e.
+    # ``device="mps"``) must select the MPS backend, not silently fall back to
+    # CPU. Previously the literal string "mps" was used as a device index,
+    # building the invalid ``torch.device("mps:mps")`` which raised and forced
+    # CPU. Only meaningful where MPS is actually available.
+    import torch
+    from cellpose import core
+
+    if not torch.backends.mps.is_available():
+        pytest.skip("MPS backend not available")
+
+    device, gpu = core.assign_device(use_torch=True, gpu=True, device="mps")
+    assert gpu
+    assert device.type == "mps"
+
+
 def itest_model_dir():
     import os, pathlib
     import numpy as np