update video depth

infer.py

@@ -80,7 +80,18 @@ def infer_pipe_video(pipe, test_image, task_name, generator, device, latents=Non
     else:
         autocast_ctx = torch.autocast(pipe.device.type)
     with autocast_ctx:
-        test_image = np.array(test_image).astype(np.float16)
+        test_image = np.array(test_image).astype(np.float32)
+        if max(test_image.shape[:2]) > 1024:
+            # resize for a maximum size of 1024
+            scale = 1024 / max(test_image.shape[:2])
+        elif min(test_image.shape[:2]) < 384:
+            # resize for a minimum size of 384
+            scale = 384 / min(test_image.shape[:2])
+        else:
+            scale = 1.0
+        new_shape = (int(test_image.shape[1] * scale), int(test_image.shape[0] * scale))
+        test_image = cv2.resize(test_image, new_shape)
+        test_image = test_image.astype(np.float16)
         test_image = torch.tensor(test_image).permute(2,0,1).unsqueeze(0)
         test_image = test_image / 127.5 - 1.0 
         test_image = test_image.to(device)