diff --git a/toolkit/captioning/caption_with_cogvlm.py b/toolkit/captioning/caption_with_cogvlm.py
index 52612645..9c1dbf23 100644
--- a/toolkit/captioning/caption_with_cogvlm.py
+++ b/toolkit/captioning/caption_with_cogvlm.py
@@ -293,6 +293,8 @@ def process_directory(
                             counter += 1
 
                         image.save(new_filepath)
+                    else:
+                        new_filepath = full_filepath
                     if args.target_backend_id:
                         upload_to_s3(s3_client, bucket_name, image, new_filename)
 
@@ -342,15 +344,23 @@ def main():
     if args.output_dir and not os.path.exists(args.output_dir):
         os.makedirs(args.output_dir)
     logger.info("Loading CogVLM model. This should only occur once.")
-    from transformers import AutoModelForCausalLM, LlamaTokenizer
+    from transformers import AutoModelForCausalLM, LlamaTokenizer, AutoTokenizer
 
-    tokenizer = LlamaTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
     logger.info(f"Loading CogVLM in {args.precision} precision.")
     if "cogvlm2" in args.model_path and torch.backends.mps.is_available():
         logger.warning(
             "Can not run CogVLM 2 on MPS because Triton is unavailable. Falling back to CogVLM 1.1"
         )
-        args.model_path = "THUDM/cogvlm-chat-hf"
+    elif "cogvlm2" in args.model_path:
+        import sysconfig
+
+        print(sysconfig.get_paths()["include"])
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.model_path, trust_remote_code=True
+        )
+    else:
+        tokenizer = LlamaTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
+
     model = AutoModelForCausalLM.from_pretrained(
         args.model_path,
         torch_dtype=torch_dtype,