diff --git a/src/utils/ultralytics_16bit_patch.py b/src/utils/ultralytics_16bit_patch.py
index 0b3e699..032f811 100644
--- a/src/utils/ultralytics_16bit_patch.py
+++ b/src/utils/ultralytics_16bit_patch.py
@@ -72,8 +72,9 @@ def apply_ultralytics_16bit_tiff_patches(*, force: bool = False) -> None:
             # logger.info(f"Loading with monkey-patched imread: {filename}")
             arr = arr.astype(np.float32)
             arr /= arr.max()
-            arr *= 2**16 - 1
-            arr = arr.astype(np.uint16)
+            arr *= 2**8 - 1
+            arr = arr.astype(np.uint8)
+            # print(arr.shape, arr.dtype, any(np.isnan(arr).flatten()), np.where(np.isnan(arr)), arr.min(), arr.max())
             return np.ascontiguousarray(arr)
 
         # logger.info(f"Loading with original imread: {filename}")
@@ -105,7 +106,7 @@ def apply_ultralytics_16bit_tiff_patches(*, force: bool = False) -> None:
     def preprocess_batch_16bit(self, batch: dict) -> dict:  # type: ignore[override]
         # Start from upstream behavior to keep device placement + multiscale identical,
         # but replace the 255 division with dtype-aware scaling.
-        logger.info(f"Preprocessing batch with monkey-patched preprocess_batch")
+        # logger.info(f"Preprocessing batch with monkey-patched preprocess_batch")
         for k, v in batch.items():
             if isinstance(v, torch.Tensor):
                 batch[k] = v.to(self.device, non_blocking=self.device.type == "cuda")