diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py
index 69ed6019a7..bb9683129b 100644
--- a/llama_cpp/llama_chat_format.py
+++ b/llama_cpp/llama_chat_format.py
@@ -1853,6 +1853,8 @@ def __init__(self, clip_model_path: str, verbose: bool = False):
                 self.clip_model_path.encode(), 0
             )
 
+        self.current_image_embed = None
+
     def __del__(self):
         with suppress_stdout_stderr(disable=self.verbose):
             if self.clip_ctx is not None and self._clip_free is not None:
@@ -1943,6 +1945,10 @@ def __call__(
                                 )
                             )
                         if content["type"] == "image_url":
+                            if self.current_image_embed is not None:
+                                with suppress_stdout_stderr(disable=self.verbose):
+                                    self._llava_cpp.llava_image_embed_free(self.current_image_embed)
+
                             image_bytes = (
                                 self.load_image(content["image_url"]["url"])
                                 if isinstance(content["image_url"], dict)
@@ -1963,6 +1969,7 @@ def __call__(
                                         len(image_bytes),
                                     )
                                 )
+                                self.current_image_embed = embed
                             try:
                                 n_past = ctypes.c_int(llama.n_tokens)
                                 n_past_p = ctypes.pointer(n_past)
@@ -1975,9 +1982,26 @@ def __call__(
                                     )
                                 assert llama.n_ctx() >= n_past.value
                                 llama.n_tokens = n_past.value
-                            finally:
+                            except:
                                 with suppress_stdout_stderr(disable=self.verbose):
                                     self._llava_cpp.llava_image_embed_free(embed)
+                        if content["type"] == "image_current":
+                            try:
+                                n_past = ctypes.c_int(llama.n_tokens)
+                                n_past_p = ctypes.pointer(n_past)
+                                with suppress_stdout_stderr(disable=self.verbose):
+                                    self._llava_cpp.llava_eval_image_embed(
+                                        ctx_llama=llama.ctx,
+                                        embed=self.current_image_embed,
+                                        n_batch=llama.n_batch,
+                                        n_past=n_past_p,
+                                    )
+                                assert llama.n_ctx() >= n_past.value
+                                llama.n_tokens = n_past.value
+                            except:
+                                with suppress_stdout_stderr(disable=self.verbose):
+                                    self._llava_cpp.llava_image_embed_free(self.current_image_embed)
+                            
             if message["role"] == "assistant" and message["content"] is not None:
                 llama.eval(
                     llama.tokenize(