Place image parts before video parts in multimodal payloads

gemini_nodes.py

@@ -1012,7 +1012,7 @@ class SSL_GeminiTextPrompt(IO.ComfyNode):
                 print(f"[ERROR] Gemini client initialization failed: {str(e)}")
                 return IO.NodeOutput(f"Gemini client initialization failed: {str(e)}", cls.generate_empty_image(), actual_seed if actual_seed is not None else 0, "", "")
 
-            # Prepare contents (video + images + prompt)
+            # Prepare contents (images + video + prompt)
             image_frames, image_batch_counts = cls._ordered_image_frames(image_inputs)
             context_image_hashes = []
             media_parts = []
@@ -1022,18 +1022,6 @@ class SSL_GeminiTextPrompt(IO.ComfyNode):
                 + len(padded_system_instruction.encode("utf-8"))
             )
 
-            if video_bytes is not None:
-                if model in cls.MEDIA_RES_MODELS and media_resolution is not None and media_resolution != "unspecified":
-                    video_part = types.Part.from_bytes(
-                        data=video_bytes,
-                        mime_type=video_mime_type,
-                        media_resolution=f"MEDIA_RESOLUTION_{media_resolution.upper()}",
-                    )
-                else:
-                    video_part = types.Part.from_bytes(data=video_bytes, mime_type=video_mime_type)
-                video_part.video_metadata = types.VideoMetadata(fps=video_fps)
-                media_parts.append(video_part)
-
             if image_frames:
                 try:
                     img_parts = []
@@ -1070,6 +1058,18 @@ class SSL_GeminiTextPrompt(IO.ComfyNode):
                 except Exception as e:
                     print(f"[ERROR] Error processing input image: {str(e)}")
                     return IO.NodeOutput(f"Error processing input image: {str(e)}", cls.generate_empty_image(), actual_seed if actual_seed is not None else 0, "", "")
+
+            if video_bytes is not None:
+                if model in cls.MEDIA_RES_MODELS and media_resolution is not None and media_resolution != "unspecified":
+                    video_part = types.Part.from_bytes(
+                        data=video_bytes,
+                        mime_type=video_mime_type,
+                        media_resolution=f"MEDIA_RESOLUTION_{media_resolution.upper()}",
+                    )
+                else:
+                    video_part = types.Part.from_bytes(data=video_bytes, mime_type=video_mime_type)
+                video_part.video_metadata = types.VideoMetadata(fps=video_fps)
+                media_parts.append(video_part)
             if media_parts:
                 contents = [
                     types.UserContent(

interactions_nodes.py

@@ -841,17 +841,6 @@ class SSL_GeminiInteractionsTextPrompt(IO.ComfyNode):
             image_frames, image_batch_counts = cls._ordered_image_frames(image_inputs)
             media_items = []
 
-            if video_bytes is not None:
-                video_stream = BytesIO(video_bytes)
-                v_item = {
-                    "type": "video",
-                    "data": video_stream,
-                    "mime_type": video_mime_type,
-                }
-                if media_resolution and media_resolution != "unspecified":
-                    v_item["resolution"] = media_resolution.lower()
-                media_items.append(v_item)
-
             if image_frames:
                 try:
                     for _, _, image_frame in image_frames:
@@ -875,12 +864,23 @@ class SSL_GeminiInteractionsTextPrompt(IO.ComfyNode):
                     )
                     print(
                         f"[INFO] Prepared Gemini image inputs: {batch_summary}, "
-                        f"total={len(media_items) - (1 if video_bytes is not None else 0)}"
+                        f"total={len(media_items)}"
                     )
                 except Exception as e:
                     print(f"[ERROR] Error processing input image: {str(e)}")
                     return IO.NodeOutput(f"Error processing input image: {str(e)}", cls.generate_empty_image(), actual_seed if actual_seed is not None else 0, "", "", "")
 
+            if video_bytes is not None:
+                video_stream = BytesIO(video_bytes)
+                v_item = {
+                    "type": "video",
+                    "data": video_stream,
+                    "mime_type": video_mime_type,
+                }
+                if media_resolution and media_resolution != "unspecified":
+                    v_item["resolution"] = media_resolution.lower()
+                media_items.append(v_item)
+
             text_item = {"type": "text", "text": padded_prompt}
             if media_items:
                 input_contents = [*media_items, text_item]

tests/test_gemini_nodes.py

@@ -531,9 +531,10 @@ def test_context_cache_reuses_full_input_after_local_fingerprint_miss(monkeypatc
     create_config = captured["creates"][0]["config"]
     assert create_config.system_instruction == "test system instruction"
     assert create_config.ttl == "900s"
-    assert create_config.contents[0].parts[0].inline_data.data == b"normalized-video"
-    assert create_config.contents[0].parts[0].inline_data.mime_type == "video/mp4"
-    assert create_config.contents[0].parts[0].video_metadata.fps == 2
+    assert create_config.contents[0].parts[0].inline_data.mime_type == "image/png"
+    assert create_config.contents[0].parts[1].inline_data.data == b"normalized-video"
+    assert create_config.contents[0].parts[1].inline_data.mime_type == "video/mp4"
+    assert create_config.contents[0].parts[1].video_metadata.fps == 2
     assert create_config.contents[0].parts[-1].text == "test prompt"
     assert len(create_config.contents[0].parts) == 3
     assert len(captured["generations"]) == 2

tests/test_interactions_nodes.py

@@ -244,6 +244,25 @@ def test_multimodal_video_input(mock_interactions_client):
     assert base64.b64decode(validated_vid.data) == video_data
 
 
+def test_multimodal_image_and_video_ordering(mock_interactions_client):
+    mock_interactions_client.responses = [{"text": "Found image and video", "id": "int_both_1"}]
+    video_data = b"fake video bytes"
+    image_tensor = torch.zeros(1, 16, 16, 3)
+    args = _execute_kwargs(
+        video={"video": FakeVideo(video_data), "fps": 2, "pad_at_start": False, "duration_aware_padding": False},
+        image_inputs={"image_1": image_tensor},
+        prompt="Describe image and video",
+    )
+    output = interactions_nodes.SSL_GeminiInteractionsTextPrompt.execute(**args)
+    assert output[0] == "Found image and video"
+    call = mock_interactions_client.calls[0]
+    items = call["input"]
+    assert len(items) == 3
+    assert items[0]["type"] == "image"
+    assert items[1]["type"] == "video"
+    assert items[2]["type"] == "text"
+
+
 def test_structured_output_with_schema(mock_interactions_client):
     schema = {
         "type": "object",