diff --git a/figures/ofa_search_cost.png b/figures/ofa_search_cost.png
new file mode 100644
index 0000000..241a124
Binary files /dev/null and b/figures/ofa_search_cost.png differ
diff --git a/figures/predictor_based_search.png b/figures/predictor_based_search.png
new file mode 100644
index 0000000..c844d3e
Binary files /dev/null and b/figures/predictor_based_search.png differ
diff --git a/figures/select_subnets.png b/figures/select_subnets.png
new file mode 100644
index 0000000..dbe7214
Binary files /dev/null and b/figures/select_subnets.png differ
diff --git a/tutorial/ofa.ipynb b/tutorial/ofa.ipynb
index 24e128c..9993b54 100644
--- a/tutorial/ofa.ipynb
+++ b/tutorial/ofa.ipynb
@@ -15,7 +15,7 @@
     "Different sub-nets can directly grab weights from the OFA network without training.\n",
     "Therefore, getting a new specialized neural network with the OFA network is highly efficient, incurring little computation cost.\n",
     "\n",
-    "![](https://hanlab.mit.edu/files/OnceForAll/figures/ofa_search_cost.png)"
+    "![](../figures/ofa_search_cost.png)"
    ]
   },
   {
@@ -300,7 +300,7 @@
    "metadata": {},
    "source": [
     "## 2. Using Pretrained Specialized OFA Sub-Networks\n",
-    "![](https://hanlab.mit.edu/files/OnceForAll/figures/select_subnets.png)\n",
+    "![](../figures/select_subnets.png)\n",
     "The specialized OFA sub-networks are \"small\" networks sampled from the \"big\" OFA network as is indicated in the figure above.\n",
     "The OFA network supports over $10^{19}$ sub-networks simultaneously, so that the deployment cost for multiple scenarios can be saved by 16$\\times$ to 1300$\\times$ under 40 deployment scenarios.\n",
     "Now, let's play with some of the sub-networks through the following interactive command line prompt (**Notice that for CPU users, this will be skipped**).\n",
@@ -372,7 +372,7 @@
     "(different from the official 50K validation set) so that we do **NOT** need to run very costly inference on ImageNet\n",
     "while searching for specialized models. Such an accuracy predictor is trained using an accuracy dataset built with the OFA network.\n",
     "\n",
-    "![](https://hanlab.mit.edu/files/OnceForAll/figures/predictor_based_search.png)"
+    "![](../figures/predictor_based_search.png)"
    ]
   },
   {
@@ -657,7 +657,7 @@
    "source": [
     "**Notice:** You can further significantly improve the accuracy of the searched sub-net by fine-tuning it on the ImageNet training set.\n",
     "Our results after fine-tuning for 25 epochs are as follows:\n",
-    "![](https://hanlab.mit.edu/files/OnceForAll/figures/diverse_hardware.png)\n",
+    "![](../figures/diverse_hardware.png)\n",
     "\n",
     "\n",
     "### 3.2 FLOPs-Constrained Efficient Deployment\n",
@@ -915,8 +915,8 @@
     "**Notice:** Again, you can further improve the accuracy of the search sub-net by fine-tuning it on ImageNet.\n",
     "The final accuracy is much better than training the same architecture from scratch.\n",
     "Our results are as follows:\n",
-    "![](https://hanlab.mit.edu/files/OnceForAll/figures/imagenet_80_acc.png)\n",
-    "![](https://hanlab.mit.edu/files/OnceForAll/figures/cnn_imagenet_new.png)\n",
+    "![](../figures/imagenet_80_acc.png)\n",
+    "![](../figures/cnn_imagenet_new.png)\n",
     "\n",
     "Congratulations! You've finished all the content of this tutorial!\n",
     "Hope you enjoy playing with the OFA Networks. If you are interested,  please refer to our paper and GitHub Repo for further details.\n",