diff --git a/utils/sstv/sstv_decoder.py b/utils/sstv/sstv_decoder.py
index dbb6591..498d52e 100644
--- a/utils/sstv/sstv_decoder.py
+++ b/utils/sstv/sstv_decoder.py
@@ -464,7 +464,13 @@ class SSTVDecoder:
                     result = vis_detector.feed(samples)
                     if result is not None:
                         vis_code, mode_name = result
-                        logger.info(f"VIS detected: code={vis_code}, mode={mode_name}")
+                        # Capture samples that arrived after the VIS STOP_BIT —
+                        # these are the start of the image and must be fed into
+                        # the image decoder before the next chunk arrives.
+                        remaining = vis_detector.remaining_buffer.copy()
+                        vis_detector.reset()
+                        logger.info(f"VIS detected: code={vis_code}, mode={mode_name}, "
+                                    f"{len(remaining)} image-start samples retained")
 
                         mode_spec = get_mode(vis_code)
                         if mode_spec:
@@ -473,6 +479,8 @@ class SSTVDecoder:
                                 mode_spec,
                                 sample_rate=SAMPLE_RATE,
                             )
+                            if len(remaining) > 0:
+                                image_decoder.feed(remaining)
                             self._emit_progress(DecodeProgress(
                                 status='decoding',
                                 mode=mode_name,
@@ -481,7 +489,6 @@ class SSTVDecoder:
                             ))
                         else:
                             logger.warning(f"No mode spec for VIS code {vis_code}")
-                            vis_detector.reset()
 
                     # Emit signal level metrics every ~500ms (every 5th 100ms chunk)
                     scope_tone: str | None = None
@@ -853,6 +860,8 @@ class SSTVDecoder:
                     result = vis_detector.feed(chunk)
                     if result is not None:
                         vis_code, mode_name = result
+                        remaining = vis_detector.remaining_buffer.copy()
+                        vis_detector.reset()
                         logger.info(f"VIS detected in file: code={vis_code}, mode={mode_name}")
 
                         mode_spec = get_mode(vis_code)
@@ -862,8 +871,8 @@ class SSTVDecoder:
                                 mode_spec,
                                 sample_rate=SAMPLE_RATE,
                             )
-                        else:
-                            vis_detector.reset()
+                            if len(remaining) > 0:
+                                image_decoder.feed(remaining)
 
         except wave.Error as e:
             logger.error(f"Error reading WAV file: {e}")
diff --git a/utils/sstv/vis.py b/utils/sstv/vis.py
index bd3d3ed..1f014e4 100644
--- a/utils/sstv/vis.py
+++ b/utils/sstv/vis.py
@@ -145,6 +145,16 @@ class VISDetector:
     def state(self) -> VISState:
         return self._state
 
+    @property
+    def remaining_buffer(self) -> np.ndarray:
+        """Unprocessed samples left after VIS detection.
+
+        Valid immediately after feed() returns a detection result and before
+        reset() is called. These samples are the start of the SSTV image and
+        must be forwarded to the image decoder.
+        """
+        return self._buffer
+
     def feed(self, samples: np.ndarray) -> tuple[int, str] | None:
         """Feed audio samples and attempt VIS detection.
 
@@ -288,9 +298,22 @@ class VISDetector:
 
             if self._tone_counter >= self._bit_windows:
                 result = self._validate_and_decode()
-                self.reset()
+                # Do NOT call reset() here. self._buffer still holds samples
+                # that arrived after the STOP_BIT window — those are the very
+                # first samples of the image. Wiping the buffer here causes all
+                # of them to be lost, making the image decoder start mid-stream
+                # and producing garbled/diagonal output.
+                # feed() will advance past the current window, leaving
+                # self._buffer pointing at the image start. The caller must
+                # read remaining_buffer and then call reset() explicitly.
+                self._state = VISState.DETECTED
                 return result
 
+        elif self._state == VISState.DETECTED:
+            # Waiting for caller to call reset() after reading remaining_buffer.
+            # Don't process any windows in this state.
+            pass
+
         return None
 
     def _decode_bit(self, samples: np.ndarray) -> int: