diff --git a/deepomatic/api/inputs.py b/deepomatic/api/inputs.py
index 5f0a338..4655ae4 100644
--- a/deepomatic/api/inputs.py
+++ b/deepomatic/api/inputs.py
@@ -25,6 +25,7 @@
 import sys
 import copy
 import base64
+import json
 
 from deepomatic.api.exceptions import DeepomaticException
 
@@ -102,6 +103,9 @@ def get_input(self):
         }
         if self.bbox is not None:
             image['bbox'] = self.bbox
+            if self.need_multipart and isinstance(image['bbox'], dict):
+                # Convert the bbox to a string to allow serialization
+                image['bbox'] = json.dumps(image['bbox'])
         if self.polygon is not None:
             image['polygon'] = self.polygon
         return {
diff --git a/tests/test_client.py b/tests/test_client.py
index d02b315..bb5e935 100644
--- a/tests/test_client.py
+++ b/tests/test_client.py
@@ -18,6 +18,7 @@
 from deepomatic.api.http_retry import HTTPRetry
 from deepomatic.api.inputs import ImageInput
 from deepomatic.api.version import __title__, __version__
+from io import BytesIO
 from requests.exceptions import ConnectionError, MissingSchema
 from tenacity import RetryError, stop_after_delay
 
@@ -259,6 +260,15 @@ def test_create_custom_reco_and_infer(self, client, custom_network):
                                    max_predictions=3)
         assert inference_schema(3, 0, 'golden retriever', 0.5) == result
 
+        # Test a binary image via multipart
+        response = requests.get(DEMO_URL)
+        img_data = BytesIO(response.content)
+        result = version.inference(inputs=[ImageInput(img_data, encoding='binary',
+                                                      bbox={"xmin": 0.2, "ymin": 0.2, "xmax": 0.8, "ymax": 0.8})],
+                                   show_discarded=True,
+                                   max_predictions=3)
+        assert inference_schema(3, 0, 'golden retriever', 0.5) == result
+
         versions = spec.versions()
         assert versions.count() > 0
         data = versions.data()