diff --git a/inference_video_parallel.py b/inference_video_parallel.py
index becdfd5..44578c9 100644
--- a/inference_video_parallel.py
+++ b/inference_video_parallel.py
@@ -6,6 +6,8 @@ import numpy as np
 from tqdm import tqdm
 from torch.nn import functional as F
 import warnings
+import _thread
+
 warnings.filterwarnings("ignore")
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@@ -48,11 +50,10 @@ else:
     
 cnt = 0
 skip_frame = 1
-
+buffer = []
 
 def write_frame(i0, infs, i1, p, user_args):
-    global skip_frame, cnt
-    
+    global skip_frame, cnt    
     for i in range(i0.shape[0]):
         # A video transition occurs.
         if p[i] > 0.2:
@@ -70,18 +71,19 @@ def write_frame(i0, infs, i1, p, user_args):
             skip_frame += 1
             continue
         
-        # Write results.
-        if user_args.png:
-            cv2.imwrite('output/{:0>7d}.png'.format(cnt), i0[i])
-            cnt += 1
-            for inf in infs:
-                cv2.imwrite('output/{:0>7d}.png'.format(cnt), inf[i])
-                cnt += 1
-        else:
-            vid_out.write(i0[i])
-            for inf in infs:
-                vid_out.write(inf[i])
+        # Write results.        
+        buffer.append(i0[i])
+        for inf in infs:
+            buffer.append(inf[i])
 
+def clear_buffer(user_args, buffer):
+    global cnt
+    for i in buffer:
+        if user_args.png:
+            cv2.imwrite('output/{:0>7d}.png'.format(cnt), i)
+            cnt += i
+        else:
+            vid_out.write(i)
 
 def make_inference(model, I0, I1, exp):
     middle = model.inference(I0, I1)
@@ -120,5 +122,9 @@ while success:
         write_frame(I0, inferences, I1, p.mean(3).mean(2).mean(1), args)
         pbar.update(4)
         img_list = img_list[-1:]
+    if len(buffer) > 100:
+        _thread.start_new_thread(clear_buffer, (args, buffer))
+        buffer = []
+_thread.start_new_thread(clear_buffer, (args, buffer))
 pbar.close()
 vid_out.release()