diff --git a/src/aare/gui/threads/prediction_subscriber.py b/src/aare/gui/threads/prediction_subscriber.py
index 6a09da19..1142435d 100644
--- a/src/aare/gui/threads/prediction_subscriber.py
+++ b/src/aare/gui/threads/prediction_subscriber.py
@@ -1,14 +1,21 @@
 import json
-from aare.common.logger_config import setup_logger
+import numpy as np
 import zmq
-from PySide6.QtCore import QThread, Signal
 
-logger = setup_logger("aareGUI")
+from PySide6.QtCore import QThread, Signal
+from PySide6.QtGui import QImage, QPixmap
+
+# If you need Bayer conversion like your SampleCameraThread did:
+import cv2
+
 
 class PredictionSubscriber(QThread):
     # emits parsed JSON payload (dict with keys: time, frame_id, shape, boxes)
     prediction = Signal(dict)
 
+    # NEW: emit the image that came with the prediction stream
+    image = Signal(QPixmap)
+
     def __init__(self, pred_zmq_url: str, topic: bytes | str = b"", parent=None):
         super().__init__(parent)
         self._ctx = zmq.Context()
@@ -22,19 +29,80 @@ class PredictionSubscriber(QThread):
         self._sock.connect(pred_zmq_url)
         self.running = True
 
+    def _try_parse_json(self, part: bytes) -> dict | None:
+        try:
+            decoded = json.loads(part.decode("utf-8"))
+            return decoded if isinstance(decoded, dict) else None
+        except Exception:
+            return None
+
+    def _decode_image(self, header: dict, data: bytes) -> QPixmap | None:
+        """
+        Supports:
+          - header["type"] == "uint8"
+          - header["shape"] == [H, W] (Bayer)  -> converted to RGB
+          - header["shape"] == [H, W, 3] (RGB) -> used directly
+        """
+        if not header or header.get("type") != "uint8":
+            return None
+        shape = header.get("shape")
+        if not shape or not isinstance(shape, (list, tuple)):
+            return None
+
+        arr = np.frombuffer(data, dtype=np.uint8)
+
+        if len(shape) == 2:
+            h, w = int(shape[0]), int(shape[1])
+            if arr.size != h * w:
+                return None
+            bayer = arr.reshape((h, w))
+            rgb = cv2.cvtColor(bayer, cv2.COLOR_BAYER_GB2RGB)
+        elif len(shape) == 3 and int(shape[2]) == 3:
+            h, w, c = int(shape[0]), int(shape[1]), int(shape[2])
+            if arr.size != h * w * c:
+                return None
+            rgb = arr.reshape((h, w, 3))
+        else:
+            return None
+
+        # If you still need the horizontal flip you had before:
+        rgb = rgb[:, ::-1, :].copy()
+
+        qimage = QImage(rgb.data, rgb.shape[1], rgb.shape[0], QImage.Format.Format_RGB888)
+        return QPixmap.fromImage(qimage)
+
     def run(self):
         while self.running:
             try:
                 parts = self._sock.recv_multipart()
                 if not parts:
                     continue
-                # publisher sends either raw JSON or [topic, json]
-                payload_bytes = parts[-1]
-                try:
-                    payload = json.loads(payload_bytes.decode("utf-8"))
-                except Exception:
-                    continue
-                self.prediction.emit(payload)
+
+                json_dicts: list[dict] = []
+                non_json_parts: list[bytes] = []
+
+                # Parse all parts; keep non-JSON as candidates for image bytes
+                for p in parts:
+                    d = self._try_parse_json(p)
+                    if d is not None:
+                        json_dicts.append(d)
+                    else:
+                        non_json_parts.append(p)
+
+                header = next((d for d in json_dicts if "shape" in d and d.get("type") == "uint8"), None)
+                detections = next((d for d in json_dicts if "boxes" in d), None)
+
+                # Heuristic: image payload is usually the largest non-JSON part
+                image_bytes = max(non_json_parts, key=len) if non_json_parts else None
+
+                if header and image_bytes:
+                    pix = self._decode_image(header, image_bytes)
+                    if pix is not None:
+                        self.image.emit(pix)
+
+                if detections:
+                    self.prediction.emit(detections)
+
             except Exception as e:
                 logger.error("PredictionSubscriber error:", e)
                 break