diff --git a/micro_sam/prompt_based_segmentation.py b/micro_sam/prompt_based_segmentation.py
index 7e38ae063..82dbfcb89 100644
--- a/micro_sam/prompt_based_segmentation.py
+++ b/micro_sam/prompt_based_segmentation.py
@@ -12,8 +12,6 @@
 
 import torch
 
-from nifty.tools import blocking
-
 from segment_anything.predictor import SamPredictor
 from segment_anything.utils.transforms import ResizeLongestSide
 
@@ -155,6 +153,7 @@ def _process_box(box, shape, original_size=None, box_extension=0):
 # and bring the points to the coordinate system of the tile.
 # Discard points that are not in the tile and warn if this happens.
 def _points_to_tile(prompts, shape, tile_shape, halo):
+    from nifty.tools import blocking
     points, labels = prompts
 
     tiling = blocking([0, 0], shape, tile_shape)
@@ -186,6 +185,7 @@ def _points_to_tile(prompts, shape, tile_shape, halo):
 
 
 def _box_to_tile(box, shape, tile_shape, halo):
+    from nifty.tools import blocking
     tiling = blocking([0, 0], shape, tile_shape)
     center = np.array([(box[0] + box[2]) / 2, (box[1] + box[3]) / 2]).round().astype("int").tolist()
     tile_id = tiling.coordinatesToBlockId(center)
@@ -205,6 +205,7 @@ def _box_to_tile(box, shape, tile_shape, halo):
 
 
 def _mask_to_tile(mask, shape, tile_shape, halo):
+    from nifty.tools import blocking
     tiling = blocking([0, 0], shape, tile_shape)
 
     coords = np.where(mask)
diff --git a/micro_sam/util.py b/micro_sam/util.py
index 531e741d0..8173d3d58 100644
--- a/micro_sam/util.py
+++ b/micro_sam/util.py
@@ -12,18 +12,14 @@
 from collections import OrderedDict
 from typing import Any, Dict, Iterable, List, Optional, Tuple, Union, Callable
 
-import elf.parallel as parallel_impl
 import imageio.v3 as imageio
 import numpy as np
 import pooch
 import segment_anything.utils.amg as amg_utils
 import torch
-import vigra
 import xxhash
 import zarr
 
-from elf.io import open_file
-from nifty.tools import blocking
 from skimage.measure import regionprops
 from skimage.segmentation import relabel_sequential
 from torchvision.ops.boxes import batched_nms
@@ -762,6 +758,7 @@ def _check_mask(tile_id):
 
 
 def _compute_tiled_features_2d(predictor, input_, tile_shape, halo, f, pbar_init, pbar_update, batch_size, mask):
+    from nifty.tools import blocking
     tiling = blocking([0, 0], input_.shape[:2], tile_shape)
     n_tiles = tiling.numberOfBlocks
 
@@ -850,6 +847,7 @@ def __next__(self):
 
 
 def _compute_tiled_features_3d(predictor, input_, tile_shape, halo, f, pbar_init, pbar_update, batch_size, mask):
+    from nifty.tools import blocking
     assert input_.ndim == 3
 
     shape = input_.shape[1:]
@@ -1301,6 +1299,7 @@ def get_centers_and_bounding_boxes(
     if mode == "p":
         center_coordinates = {prop.label: prop.centroid for prop in properties}
     elif mode == "v":
+        import vigra
         center_coordinates = vigra.filters.eccentricityCenters(segmentation.astype('float32'))
         center_coordinates = {i: coord for i, coord in enumerate(center_coordinates) if i > 0}
 
@@ -1324,6 +1323,7 @@ def load_image_data(path: str, key: Optional[str] = None, lazy_loading: bool = F
     if key is None:
         image_data = imageio.imread(path)
     else:
+        from elf.io import open_file
         with open_file(path, mode="r") as f:
             image_data = f[key]
             if not lazy_loading:
@@ -1807,6 +1807,7 @@ def require_numpy(mask):
             segmentation[this_mask] = this_seg_id
         seg_id = this_seg_id + 1
 
+    import elf.parallel as parallel_impl
     block_shape = (512, 512)
     if label_masks:
         segmentation_cc = np.zeros_like(segmentation, dtype=segmentation.dtype)
diff --git a/setup.cfg b/setup.cfg
index 24bbfec82..64614b6e7 100644
--- a/setup.cfg
+++ b/setup.cfg
@@ -31,6 +31,18 @@ project_urls =
 [options]
 packages = find:
 python_requires = >=3.10
+install_requires =
+    imageio
+    numpy
+    pooch
+    scikit-image
+    scipy
+    segment-anything
+    torch>=2.5
+    torchvision
+    tqdm
+    xxhash
+    zarr
 include_package_data = True
 package_dir =
     = .