diff --git a/editor/training/histogram_dataset.py b/editor/training/histogram_dataset.py
index 2693d20..7d44e4e 100644
--- a/editor/training/histogram_dataset.py
+++ b/editor/training/histogram_dataset.py
@@ -1,5 +1,5 @@
 from torch.utils.data import Dataset
-from typing import List
+from typing import List, Optional
 from editor.utils import compute_histogram
 from .random_edit import random_edit
 from PIL import Image
@@ -20,11 +20,13 @@ class HistogramDataset(Dataset):
         bin_count: int = 32,
         target_size=(480, 480),
         delete_corrupt_images: bool = False,
+        cache_path: Optional[Path] = None,
     ):
         self._paths = sorted(paths)
         self._edit_count = edit_count
         self._bin_count = bin_count
         self._target_size = target_size
+        self._cache_path = cache_path
 
         if delete_corrupt_images:
             self._delete_corrupt_images()
@@ -44,6 +46,14 @@ class HistogramDataset(Dataset):
         return len(self._paths) * self._edit_count
 
     def __getitem__(self, idx):
+        if self._cache_path is not None:
+            self._cached_data_path = self._cache_path / f"{idx}.pt"
+            if self._cached_data_path.exists():
+                try:
+                    return torch.load(self._cached_data_path)
+                except:
+                    print(f"Failed to load {self._cached_data_path}, regenerating...")
+
         original_idx = idx // self._edit_count
         original_path = self._paths[original_idx]
         original = Image.open(original_path)
@@ -58,7 +68,12 @@ class HistogramDataset(Dataset):
             edited, bins=self._bin_count, normalize=True
         )
 
-        return (
+        result = (
             torch.tensor(edited_histogram, dtype=torch.float).unsqueeze(0),
             torch.tensor(original_histogram, dtype=torch.float).unsqueeze(0),
         )
+
+        if self._cache_path is not None:
+            torch.save(result, self._cached_data_path)
+
+        return result