diff --git a/docs/anime_video_model.md b/docs/anime_video_model.md
index ba211ac..0b0ce69 100644
--- a/docs/anime_video_model.md
+++ b/docs/anime_video_model.md
@@ -36,7 +36,12 @@ The following are some demos (best view in the full screen mode).
 # download model
 wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth -P realesrgan/weights
 # inference
-python inference_realesrgan_video.py -i inputs/video/onepiece_demo.mp4 -n realesr-animevideov3 -s 2 --suffix outx2
+python inference_realesrgan_video.py -i inputs/video/onepiece_demo.mp4 -n realesr-animevideov3 -s 2 --suffix outx2 --stream
+```
+```console
+Usage:
+--stream                 with this option, the enhanced frames are sent directly to a ffmpeg stream,
+                         avoiding storing large (usually tens of GB) intermediate results.        
 ```
 
 ### NCNN Executable File
diff --git a/inference_realesrgan_video.py b/inference_realesrgan_video.py
index 3008397..d0da9fd 100644
--- a/inference_realesrgan_video.py
+++ b/inference_realesrgan_video.py
@@ -1,6 +1,8 @@
 import argparse
+import cv2
 import glob
 import mimetypes
+import numpy as np
 import os
 import queue
 import shutil
@@ -13,112 +15,129 @@ from realesrgan import IOConsumer, PrefetchReader, RealESRGANer
 from realesrgan.archs.srvgg_arch import SRVGGNetCompact
 
 
-def main():
-    """Inference demo for Real-ESRGAN.
-    It mainly for restoring anime videos.
-
-    """
-    parser = argparse.ArgumentParser()
-    parser.add_argument('-i', '--input', type=str, default='inputs', help='Input video, image or folder')
-    parser.add_argument(
-        '-n',
-        '--model_name',
-        type=str,
-        default='realesr-animevideov3',
-        help=('Model names: realesr-animevideov3 | RealESRGAN_x4plus_anime_6B | RealESRGAN_x4plus | RealESRNet_x4plus |'
-              ' RealESRGAN_x2plus | '
-              'Default:realesr-animevideov3'))
-    parser.add_argument('-o', '--output', type=str, default='results', help='Output folder')
-    parser.add_argument('-s', '--outscale', type=float, default=4, help='The final upsampling scale of the image')
-    parser.add_argument('--suffix', type=str, default='out', help='Suffix of the restored video')
-    parser.add_argument('-t', '--tile', type=int, default=0, help='Tile size, 0 for no tile during testing')
-    parser.add_argument('--tile_pad', type=int, default=10, help='Tile padding')
-    parser.add_argument('--pre_pad', type=int, default=0, help='Pre padding size at each border')
-    parser.add_argument('--face_enhance', action='store_true', help='Use GFPGAN to enhance face')
-    parser.add_argument(
-        '--fp32', action='store_true', help='Use fp32 precision during inference. Default: fp16 (half precision).')
-    parser.add_argument('--fps', type=float, default=None, help='FPS of the output video')
-    parser.add_argument('--consumer', type=int, default=4, help='Number of IO consumers')
-
-    parser.add_argument(
-        '--alpha_upsampler',
-        type=str,
-        default='realesrgan',
-        help='The upsampler for the alpha channels. Options: realesrgan | bicubic')
-    parser.add_argument(
-        '--ext',
-        type=str,
-        default='auto',
-        help='Image extension. Options: auto | jpg | png, auto means using the same extension as inputs')
-    args = parser.parse_args()
-
-    # ---------------------- determine models according to model names ---------------------- #
-    args.model_name = args.model_name.split('.pth')[0]
-    if args.model_name in ['RealESRGAN_x4plus', 'RealESRNet_x4plus']:  # x4 RRDBNet model
-        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
-        netscale = 4
-    elif args.model_name in ['RealESRGAN_x4plus_anime_6B']:  # x4 RRDBNet model with 6 blocks
-        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4)
-        netscale = 4
-    elif args.model_name in ['RealESRGAN_x2plus']:  # x2 RRDBNet model
-        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
-        netscale = 2
-    elif args.model_name in ['realesr-animevideov3']:  # x4 VGG-style model (XS size)
-        model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu')
-        netscale = 4
-
-    # ---------------------- determine model paths ---------------------- #
-    model_path = os.path.join('experiments/pretrained_models', args.model_name + '.pth')
-    if not os.path.isfile(model_path):
-        model_path = os.path.join('realesrgan/weights', args.model_name + '.pth')
-    if not os.path.isfile(model_path):
-        raise ValueError(f'Model {args.model_name} does not exist.')
-
-    # restorer
-    upsampler = RealESRGANer(
-        scale=netscale,
-        model_path=model_path,
-        model=model,
-        tile=args.tile,
-        tile_pad=args.tile_pad,
-        pre_pad=args.pre_pad,
-        half=not args.fp32)
-
-    if args.face_enhance:  # Use GFPGAN for face enhancement
-        from gfpgan import GFPGANer
-        face_enhancer = GFPGANer(
-            model_path='https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth',
-            upscale=args.outscale,
-            arch='clean',
-            channel_multiplier=2,
-            bg_upsampler=upsampler)
-    os.makedirs(args.output, exist_ok=True)
-    # for saving restored frames
-    save_frame_folder = os.path.join(args.output, 'frames_tmpout')
-    os.makedirs(save_frame_folder, exist_ok=True)
-
+def get_frames(args, extract_frames=False):
     # input can be a video file / a folder of frames / an image
+    is_video = False
     if mimetypes.guess_type(args.input)[0].startswith('video'):  # is a video file
+        is_video = True
         video_name = os.path.splitext(os.path.basename(args.input))[0]
-        frame_folder = os.path.join('tmp_frames', video_name)
-        os.makedirs(frame_folder, exist_ok=True)
-        # use ffmpeg to extract frames
-        os.system(f'ffmpeg -i {args.input} -qscale:v 1 -qmin 1 -qmax 1 -vsync 0  {frame_folder}/frame%08d.png')
-        # get image path list
-        paths = sorted(glob.glob(os.path.join(frame_folder, '*')))
+        if extract_frames:
+            frame_folder = os.path.join('tmp_frames', video_name)
+            os.makedirs(frame_folder, exist_ok=True)
+            # use ffmpeg to extract frames
+            os.system(f'ffmpeg -i {args.input} -qscale:v 1 -qmin 1 -qmax 1 -vsync 0  {frame_folder}/frame%08d.png')
+            # get image path list
+            paths = sorted(glob.glob(os.path.join(frame_folder, '*')))
+        else:
+            paths = []
         # get input video fps
         if args.fps is None:
-
             import ffmpeg
             probe = ffmpeg.probe(args.input)
             video_streams = [stream for stream in probe['streams'] if stream['codec_type'] == 'video']
             args.fps = eval(video_streams[0]['avg_frame_rate'])
     elif mimetypes.guess_type(args.input)[0].startswith('image'):  # is an image file
         paths = [args.input]
-        video_name = 'video'
     else:
         paths = sorted(glob.glob(os.path.join(args.input, '*')))
-        video_name = 'video'
+        assert len(paths) > 0, 'the input folder is empty'
+
+    if args.fps is None:
+        args.fps = 24
+
+    return is_video, paths
+
+
+def inference_stream(args, upsampler, face_enhancer):
+    try:
+        import ffmpeg
+    except ImportError:
+        import pip
+        pip.main(['install', '--user', 'ffmpeg-python'])
+        import ffmpeg
+
+    is_video, paths = get_frames(args, extract_frames=False)
+    video_name = os.path.splitext(os.path.basename(args.input))[0]
+    video_save_path = os.path.join(args.output, f'{video_name}_{args.suffix}.mp4')
+
+    # decoder
+    if is_video:
+        # get height and width
+        probe = ffmpeg.probe(args.input)
+        video_streams = [stream for stream in probe['streams'] if stream['codec_type'] == 'video']
+        width = video_streams[0]['width']
+        height = video_streams[0]['height']
+
+        # set up frame decoder
+        decoder = (
+            ffmpeg.input(args.input).output('pipe:', format='rawvideo', pix_fmt='rgb24', loglevel='warning').run_async(
+                pipe_stdin=True, pipe_stdout=True, cmd=args.ffmpeg_bin))
+    else:
+        from PIL import Image
+        tmp_img = Image.open(paths[0])
+        width, height = tmp_img.size
+        idx = 0
+
+    out_width, out_height = int(width * args.outscale), int(height * args.outscale)
+    if out_height > 2160:
+        print('You are generating video that is larger than 4K, which will be very slow due to IO speed.',
+              'We highly recommend to decrease the outscale(aka, -s).')
+    # encoder
+    if is_video:
+        audio = ffmpeg.input(args.input).audio
+        encoder = (
+            ffmpeg.input(
+                'pipe:', format='rawvideo', pix_fmt='rgb24', s=f'{out_width}x{out_height}', framerate=args.fps).output(
+                    audio, video_save_path, pix_fmt='yuv420p', vcodec='libx264', loglevel='info',
+                    acodec='copy').overwrite_output().run_async(pipe_stdin=True, pipe_stdout=True, cmd=args.ffmpeg_bin))
+    else:
+        encoder = (
+            ffmpeg.input(
+                'pipe:', format='rawvideo', pix_fmt='rgb24', s=f'{out_width}x{out_height}',
+                framerate=args.fps).output(video_save_path, pix_fmt='yuv420p', vcodec='libx264',
+                                           loglevel='info').overwrite_output().run_async(
+                                               pipe_stdin=True, pipe_stdout=True, cmd=args.ffmpeg_bin))
+
+    while True:
+        if is_video:
+            img_bytes = decoder.stdout.read(width * height * 3)  # 3 bytes for one pixel
+            if not img_bytes:
+                break
+            img = np.frombuffer(img_bytes, np.uint8).reshape([height, width, 3])
+        else:
+            if idx >= len(paths):
+                break
+            img = cv2.imread(paths[idx])
+            idx += 1
+
+        try:
+            if args.face_enhance:
+                _, _, output = face_enhancer.enhance(img, has_aligned=False, only_center_face=False, paste_back=True)
+            else:
+                output, _ = upsampler.enhance(img, outscale=args.outscale)
+        except RuntimeError as error:
+            print('Error', error)
+            print('If you encounter CUDA out of memory, try to set --tile with a smaller number.')
+        else:
+            output = output.astype(np.uint8).tobytes()
+            encoder.stdin.write(output)
+
+        torch.cuda.synchronize()
+
+    if is_video:
+        decoder.stdin.close()
+        decoder.wait()
+    encoder.stdin.close()
+    encoder.wait()
+
+
+def inference_frames(args, upsampler, face_enhancer):
+    is_video, paths = get_frames(args, extract_frames=True)
+    video_name = os.path.splitext(os.path.basename(args.input))[0]
+
+    # for saving restored frames
+    save_frame_folder = os.path.join(args.output, video_name, 'frames_tmpout')
+    os.makedirs(save_frame_folder, exist_ok=True)
 
     timer = AvgTimer()
     timer.start()
@@ -177,9 +196,109 @@ def main():
               f' -map 0:v:0 -map 1:a:0 -c:a copy -c:v libx264 -r {args.fps} -pix_fmt yuv420p  {video_save_path}')
     # delete tmp file
     shutil.rmtree(save_frame_folder)
+    frame_folder = os.path.join('tmp_frames', video_name)
     if os.path.isdir(frame_folder):
         shutil.rmtree(frame_folder)
 
 
+def main():
+    """Inference demo for Real-ESRGAN.
+    It mainly for restoring anime videos.
+
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-i', '--input', type=str, default='inputs', help='Input video, image or folder')
+    parser.add_argument(
+        '-n',
+        '--model_name',
+        type=str,
+        default='realesr-animevideov3',
+        help=('Model names: realesr-animevideov3 | RealESRGAN_x4plus_anime_6B | RealESRGAN_x4plus | RealESRNet_x4plus |'
+              ' RealESRGAN_x2plus | '
+              'Default:realesr-animevideov3'))
+    parser.add_argument('-o', '--output', type=str, default='results', help='Output folder')
+    parser.add_argument('-s', '--outscale', type=float, default=4, help='The final upsampling scale of the image')
+    parser.add_argument('--suffix', type=str, default='out', help='Suffix of the restored video')
+    parser.add_argument('-t', '--tile', type=int, default=0, help='Tile size, 0 for no tile during testing')
+    parser.add_argument('--tile_pad', type=int, default=10, help='Tile padding')
+    parser.add_argument('--pre_pad', type=int, default=0, help='Pre padding size at each border')
+    parser.add_argument('--face_enhance', action='store_true', help='Use GFPGAN to enhance face')
+    parser.add_argument(
+        '--fp32', action='store_true', help='Use fp32 precision during inference. Default: fp16 (half precision).')
+    parser.add_argument('--fps', type=float, default=None, help='FPS of the output video')
+    parser.add_argument('--consumer', type=int, default=4, help='Number of IO consumers')
+    parser.add_argument('--stream', action='store_true')
+    parser.add_argument('--ffmpeg_bin', type=str, default='ffmpeg', help='The path to ffmpeg')
+
+    parser.add_argument(
+        '--alpha_upsampler',
+        type=str,
+        default='realesrgan',
+        help='The upsampler for the alpha channels. Options: realesrgan | bicubic')
+    parser.add_argument(
+        '--ext',
+        type=str,
+        default='auto',
+        help='Image extension. Options: auto | jpg | png, auto means using the same extension as inputs')
+    args = parser.parse_args()
+
+    args.input = args.input.rstrip('/').rstrip('\\')
+
+    # ---------------------- determine models according to model names ---------------------- #
+    args.model_name = args.model_name.split('.pth')[0]
+    if args.model_name in ['RealESRGAN_x4plus', 'RealESRNet_x4plus']:  # x4 RRDBNet model
+        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
+        netscale = 4
+    elif args.model_name in ['RealESRGAN_x4plus_anime_6B']:  # x4 RRDBNet model with 6 blocks
+        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4)
+        netscale = 4
+    elif args.model_name in ['RealESRGAN_x2plus']:  # x2 RRDBNet model
+        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
+        netscale = 2
+    elif args.model_name in ['realesr-animevideov3']:  # x4 VGG-style model (XS size)
+        model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu')
+        netscale = 4
+
+    # ---------------------- determine model paths ---------------------- #
+    model_path = os.path.join('experiments/pretrained_models', args.model_name + '.pth')
+    if not os.path.isfile(model_path):
+        model_path = os.path.join('realesrgan/weights', args.model_name + '.pth')
+    if not os.path.isfile(model_path):
+        raise ValueError(f'Model {args.model_name} does not exist.')
+
+    # restorer
+    upsampler = RealESRGANer(
+        scale=netscale,
+        model_path=model_path,
+        model=model,
+        tile=args.tile,
+        tile_pad=args.tile_pad,
+        pre_pad=args.pre_pad,
+        half=not args.fp32)
+
+    if 'anime' in args.model_name and args.face_enhance:
+        print('face_enhance is not supported in anime models, we turned this option off for you. '
+              'if you insist on turning it on, please manually comment the relevant lines of code.')
+        args.face_enhance = False
+
+    if args.face_enhance:  # Use GFPGAN for face enhancement
+        from gfpgan import GFPGANer
+        face_enhancer = GFPGANer(
+            model_path='https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth',
+            upscale=args.outscale,
+            arch='clean',
+            channel_multiplier=2,
+            bg_upsampler=upsampler)
+    else:
+        face_enhancer = None
+
+    os.makedirs(args.output, exist_ok=True)
+
+    if args.stream:
+        inference_stream(args, upsampler, face_enhancer)
+    else:
+        inference_frames(args, upsampler, face_enhancer)
+
+
 if __name__ == '__main__':
     main()