diff --git a/README_zh.md b/README_zh.md
index a68630ea..13379561 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -41,7 +41,7 @@ ModelScope开源了数百个(当前600+)模型，涵盖自然语言处理、计
 
 <p align="center">
     <br>
-    <img src="https://modelscope.oss-cn-beijing.aliyuncs.com/resource/inference.gif" width="1024"/>
+    <img src="data/resource/inference.gif" width="1024"/>
     <br>
 <p>
 
@@ -129,7 +129,7 @@ ModelScope开源了数百个(当前600+)模型，涵盖自然语言处理、计
 
 给定一张图片，你可以使用如下代码进行人像抠图.
 
-![image](https://resouces.modelscope.cn/document/docdata/2023-2-16_20:53/dist/ModelScope%20Library%E6%95%99%E7%A8%8B/resources/1656989748829-9ab3aa9b-461d-44f8-98fb-c85bc6f670f9.png)
+![image](data/resource/portrait_input.png)
 
 ```python
 >>> import cv2
@@ -140,7 +140,7 @@ ModelScope开源了数百个(当前600+)模型，涵盖自然语言处理、计
 >>> cv2.imwrite('result.png', result['output_img'])
 ```
 输出图像如下
-![image](https://resouces.modelscope.cn/document/docdata/2023-2-16_20:53/dist/ModelScope%20Library%E6%95%99%E7%A8%8B/resources/1656989768092-5470f8ac-cda8-4703-ac98-dbb6fd675b34.png)
+![image](data/resource/portrait_output.png)
 
 对于微调和评估模型， 你需要通过十多行代码构建dataset和trainer，调用`trainer.train()`和`trainer.evaluate()`即可。
 
diff --git a/data/resource/inference.gif b/data/resource/inference.gif
new file mode 100644
index 00000000..81e58a85
Binary files /dev/null and b/data/resource/inference.gif differ
diff --git a/data/resource/portrait_input.png b/data/resource/portrait_input.png
new file mode 100644
index 00000000..de3f1918
--- /dev/null
+++ b/data/resource/portrait_input.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af83a94899a6d23339c3ecc5c4c58c57c835af57b531a2f4c50461184f820141
+size 603621
diff --git a/data/resource/portrait_output.png b/data/resource/portrait_output.png
new file mode 100644
index 00000000..d2ecc916
--- /dev/null
+++ b/data/resource/portrait_output.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28f6d784547c295711f1a8b7c83cf7e8d19b6361de56e9a69667fc9c9b8a429a
+size 661491