diff --git a/tools/camerastream/compressed_vipc.py b/tools/camerastream/compressed_vipc.py
index e8fc27236a..040c2e18f0 100755
--- a/tools/camerastream/compressed_vipc.py
+++ b/tools/camerastream/compressed_vipc.py
@@ -7,6 +7,8 @@ import multiprocessing
 from cereal.visionipc.visionipc_pyx import VisionIpcServer, VisionStreamType  # pylint: disable=no-name-in-module, import-error
 W, H = 1928, 1208
 
+V4L2_BUF_FLAG_KEYFRAME = 8
+
 def writer(fn, addr, sock_name):
   import cereal.messaging as messaging
   HEADER = b"\x00\x00\x00\x01\x40\x01\x0c\x01\xff\xff\x01\x60\x00\x00\x03\x00\xb0\x00\x00\x03\x00\x00\x03\x00\x96\xac\x09\x00\x00\x00\x01\x42\x01\x01\x01\x60\x00\x00\x03\x00\xb0\x00\x00\x03\x00\x00\x03\x00\x96\xa0\x03\xd0\x80\x13\x07\x1b\x2e\x5a\xee\x4c\x92\xea\x00\xbb\x42\x84\xa0\x00\x00\x00\x01\x44\x01\xc0\xe2\x4f\x09\xc1\x80\xc6\x08\x40\x00"
@@ -29,7 +31,7 @@ def writer(fn, addr, sock_name):
       if evta.idx != 0 and evta.idx != (last_idx+1):
         print("DROP!")
       last_idx = evta.idx
-      if len(evta.data) > 4 and evta.data[4] == 0x26:
+      if evta.flags & V4L2_BUF_FLAG_KEYFRAME or evta.flags == 0x7f001030:
         seen_iframe = True
       if not seen_iframe:
         print("waiting for iframe")
@@ -37,15 +39,16 @@ def writer(fn, addr, sock_name):
       fifo_file.write(evta.data)
       fifo_file.flush()
 
-def decoder_nvidia(fn, vipc_server, vst, yuv=False):
+def decoder_nvidia(fn, vipc_server, vst, yuv=True, rgb=False):
   sys.path.append("/raid.dell2/PyNvCodec")
   import PyNvCodec as nvc # pylint: disable=import-error
   decoder = nvc.PyNvDecoder(fn, 0, {"probesize": "32"})
   cc1 = nvc.ColorspaceConversionContext(nvc.ColorSpace.BT_709, nvc.ColorRange.JPEG)
 
-  conv = nvc.PySurfaceConverter(W, H, nvc.PixelFormat.NV12, nvc.PixelFormat.BGR, 0)
-  nvDwn = nvc.PySurfaceDownloader(W, H, nvc.PixelFormat.BGR, 0)
-  img = np.ndarray((H,W,3), dtype=np.uint8)
+  if rgb:
+    conv = nvc.PySurfaceConverter(W, H, nvc.PixelFormat.NV12, nvc.PixelFormat.BGR, 0)
+    nvDwn = nvc.PySurfaceDownloader(W, H, nvc.PixelFormat.BGR, 0)
+    img = np.ndarray((H,W,3), dtype=np.uint8)
 
   if yuv:
     conv_yuv = nvc.PySurfaceConverter(W, H, nvc.PixelFormat.NV12, nvc.PixelFormat.YUV420, 0)
@@ -57,22 +60,24 @@ def decoder_nvidia(fn, vipc_server, vst, yuv=False):
     rawSurface = decoder.DecodeSingleSurface()
     if rawSurface.Empty():
       continue
-    convSurface = conv.Execute(rawSurface, cc1)
-    nvDwn.DownloadSingleSurface(convSurface, img)
-    vipc_server.send(vst, img.flatten().data, cnt, 0, 0)
+    if rgb:
+      convSurface = conv.Execute(rawSurface, cc1)
+      nvDwn.DownloadSingleSurface(convSurface, img)
+      vipc_server.send(vst, img.flatten().data, cnt, 0, 0)
     if yuv:
       convSurface = conv_yuv.Execute(rawSurface, cc1)
       nvDwn_yuv.DownloadSingleSurface(convSurface, img_yuv)
       vipc_server.send(vst+3, img_yuv.flatten().data, cnt, 0, 0)
     cnt += 1
 
-def decoder_ffmpeg(fn, vipc_server, vst, yuv=False):
+def decoder_ffmpeg(fn, vipc_server, vst, yuv=True, rgb=False):
   import av # pylint: disable=import-error
   container = av.open(fn, options={"probesize": "32"})
   cnt = 0
   for frame in container.decode(video=0):
-    img = frame.to_ndarray(format=av.video.format.VideoFormat('bgr24'))
-    vipc_server.send(vst, img.flatten().data, cnt, 0, 0)
+    if rgb:
+      img = frame.to_ndarray(format=av.video.format.VideoFormat('bgr24'))
+      vipc_server.send(vst, img.flatten().data, cnt, 0, 0)
     if yuv:
       img_yuv = frame.to_ndarray(format=av.video.format.VideoFormat('yuv420p'))
       vipc_server.send(vst+3, img_yuv.flatten().data, cnt, 0, 0)
@@ -83,7 +88,7 @@ if __name__ == "__main__":
   parser = argparse.ArgumentParser(description='Decode video streams and broacast on VisionIPC')
   parser.add_argument("addr", help="Address of comma 3")
   parser.add_argument('--nvidia', action='store_true', help='Use nvidia instead of ffmpeg')
-  parser.add_argument('--yuv', action='store_true', help='Also broadcast YUV')
+  parser.add_argument('--rgb', action='store_true', help='Also broadcast RGB')
   parser.add_argument("--cams", default="0,1,2", help="Cameras to decode")
   args = parser.parse_args()
 
@@ -96,9 +101,9 @@ if __name__ == "__main__":
 
   vipc_server = VisionIpcServer("camerad")
   for vst in cams.values():
-    vipc_server.create_buffers(vst, 4, True, W, H)
-    if args.yuv:
-      vipc_server.create_buffers(vst+3, 4, False, W, H)
+    if args.rgb:
+      vipc_server.create_buffers(vst, 4, True, W, H)
+    vipc_server.create_buffers(vst+3, 4, False, W, H)
   vipc_server.start_listener()
 
   for k,v in cams.items():
@@ -108,6 +113,6 @@ if __name__ == "__main__":
     os.mkfifo(FIFO_NAME)
     multiprocessing.Process(target=writer, args=(FIFO_NAME, sys.argv[1], k)).start()
     if args.nvidia:
-      multiprocessing.Process(target=decoder_nvidia, args=(FIFO_NAME, vipc_server, v, args.yuv)).start()
+      multiprocessing.Process(target=decoder_nvidia, args=(FIFO_NAME, vipc_server, v, True, args.rgb)).start()
     else:
-      multiprocessing.Process(target=decoder_ffmpeg, args=(FIFO_NAME, vipc_server, v, args.yuv)).start()
+      multiprocessing.Process(target=decoder_ffmpeg, args=(FIFO_NAME, vipc_server, v, True, args.rgb)).start()