diff --git a/frontend/src/pages/UploadPage/UploadPage.tsx b/frontend/src/pages/UploadPage/UploadPage.tsx
index a6a95e15..f71b526d 100644
--- a/frontend/src/pages/UploadPage/UploadPage.tsx
+++ b/frontend/src/pages/UploadPage/UploadPage.tsx
@@ -481,10 +481,15 @@ export default function UploadPage() {
        }
      });
 
-    const modelName = localStorage.getItem(SELECTED_MODEL_KEY);
-    if (modelName) {
-      fd.append('model_name', modelName);
-    }
+     // Use manual mode if selected, otherwise use the selected model from localStorage
+     if (isManualMode) {
+       fd.append('model_name', 'manual');
+     } else {
+       const modelName = localStorage.getItem(SELECTED_MODEL_KEY);
+       if (modelName) {
+         fd.append('model_name', modelName);
+       }
+     }
 
      const mapRes = await fetch('/api/images/multi', { method: 'POST', body: fd });
       const mapJson = await readJsonSafely(mapRes);
diff --git a/py_backend/app/routers/upload.py b/py_backend/app/routers/upload.py
index bee004fb..3a01f4ae 100644
--- a/py_backend/app/routers/upload.py
+++ b/py_backend/app/routers/upload.py
@@ -740,19 +740,28 @@ async def upload_multiple_images(
                 "fallback_reason": result.get("fallback_reason")
             })
         
-        # Update individual image metadata if VLM provided it
-        metadata_images = metadata.get("metadata_images", {})
-        if metadata_images and isinstance(metadata_images, dict):
-            for i, img in enumerate(uploaded_images):
-                image_key = f"image{i+1}"
-                if image_key in metadata_images:
-                    img_metadata = metadata_images[image_key]
-                    if isinstance(img_metadata, dict):
-                        # Update image with individual metadata
-                        img.source = img_metadata.get("source", img.source)
-                        img.event_type = img_metadata.get("type", img.event_type)
-                        img.epsg = img_metadata.get("epsg", img.epsg)
-                        img.countries = img_metadata.get("countries", img.countries)
+        # Update individual image metadata if VLM provided it (skip for manual mode)
+        if actual_model and actual_model.lower() != "manual":
+            metadata_images = metadata.get("metadata_images", {})
+            if metadata_images and isinstance(metadata_images, dict):
+                for i, img in enumerate(uploaded_images):
+                    image_key = f"image{i+1}"
+                    if image_key in metadata_images:
+                        img_metadata = metadata_images[image_key]
+                        if isinstance(img_metadata, dict):
+                            # Update image with individual metadata only if values are not empty
+                            vlm_source = img_metadata.get("source", "")
+                            if vlm_source and vlm_source.strip():
+                                img.source = vlm_source
+                            vlm_event_type = img_metadata.get("type", "")
+                            if vlm_event_type and vlm_event_type.strip():
+                                img.event_type = vlm_event_type
+                            vlm_epsg = img_metadata.get("epsg", "")
+                            if vlm_epsg and vlm_epsg.strip():
+                                img.epsg = vlm_epsg
+                            vlm_countries = img_metadata.get("countries", [])
+                            if vlm_countries:
+                                img.countries = vlm_countries
         
         # Ensure we never use 'random' as the model name in the database
         final_model_name = actual_model if actual_model != "random" else "STUB_MODEL"
@@ -778,6 +787,8 @@ async def upload_multiple_images(
         
     except Exception as e:
         logger.debug(f"VLM error: {e}")
+        # Rollback any pending changes before creating fallback caption
+        db.rollback()
         # Create fallback caption
         fallback_text = f"Analysis of {len(image_bytes_list)} images"
         caption = crud.create_caption(