diff --git a/PaddleCV/PaddleGAN/trainer/AttGAN.py b/PaddleCV/PaddleGAN/trainer/AttGAN.py
index 41a0fcd2a64c15a260644360516c66a43bbe931a..d7d84efdc3c388fb28e86bef971cb38da76198c5 100644
--- a/PaddleCV/PaddleGAN/trainer/AttGAN.py
+++ b/PaddleCV/PaddleGAN/trainer/AttGAN.py
@@ -165,9 +165,10 @@ class DTrainer():
         grad_shape = grad.shape
         grad = fluid.layers.reshape(
             grad, [-1, grad_shape[1] * grad_shape[2] * grad_shape[3]])
+        epsilon = 1e-16
         norm = fluid.layers.sqrt(
             fluid.layers.reduce_sum(
-                fluid.layers.square(grad), dim=1))
+                fluid.layers.square(grad), dim=1) + epsilon)
         gp = fluid.layers.reduce_mean(fluid.layers.square(norm - 1.0))
         return gp
 
diff --git a/PaddleCV/PaddleGAN/trainer/STGAN.py b/PaddleCV/PaddleGAN/trainer/STGAN.py
index aeab44e14872298a0b374bef8ee0936d96137947..7b3c88a80207f7b5e9cb6304e86e75f1a7de7839 100644
--- a/PaddleCV/PaddleGAN/trainer/STGAN.py
+++ b/PaddleCV/PaddleGAN/trainer/STGAN.py
@@ -163,9 +163,10 @@ class DTrainer():
         grad_shape = grad.shape
         grad = fluid.layers.reshape(
             grad, [-1, grad_shape[1] * grad_shape[2] * grad_shape[3]])
+        epsilon = 1e-16
         norm = fluid.layers.sqrt(
             fluid.layers.reduce_sum(
-                fluid.layers.square(grad), dim=1))
+                fluid.layers.square(grad), dim=1) + epsilon)
         gp = fluid.layers.reduce_mean(fluid.layers.square(norm - 1.0))
         return gp
 
diff --git a/PaddleCV/PaddleGAN/trainer/StarGAN.py b/PaddleCV/PaddleGAN/trainer/StarGAN.py
index 6aef1fb4d796511e67a63f696f2502078ce05e73..c71b1bf7d75672dbc40202ef4de1fb48a8179d4f 100644
--- a/PaddleCV/PaddleGAN/trainer/StarGAN.py
+++ b/PaddleCV/PaddleGAN/trainer/StarGAN.py
@@ -189,9 +189,10 @@ class DTrainer():
         grad_shape = grad.shape
         grad = fluid.layers.reshape(
             grad, [-1, grad_shape[1] * grad_shape[2] * grad_shape[3]])
+        epsilon = 1e-16
         norm = fluid.layers.sqrt(
             fluid.layers.reduce_sum(
-                fluid.layers.square(grad), dim=1))
+                fluid.layers.square(grad), dim=1) + epsilon)
         gp = fluid.layers.reduce_mean(fluid.layers.square(norm - 1.0))
         return gp
 
diff --git a/PaddleCV/PaddleGAN/util/utility.py b/PaddleCV/PaddleGAN/util/utility.py
index 3c624d090801b60095bc4f60d9436bb0304d6900..0f4f216e392330c80f4164ebec233e18768428e0 100644
--- a/PaddleCV/PaddleGAN/util/utility.py
+++ b/PaddleCV/PaddleGAN/util/utility.py
@@ -113,7 +113,7 @@ def save_test_image(epoch,
             images = [real_img_temp]
             for i in range(cfg.c_dim):
                 label_trg_tmp = copy.deepcopy(label_org)
-                label_trg_tmp[0][i] = 1.0 - label_trg_tmp[j][i]
+                label_trg_tmp[0][i] = 1.0 - label_trg_tmp[0][i]
                 label_trg = check_attribute_conflict(
                     label_trg_tmp, attr_names[i], attr_names)
                 tensor_label_trg = fluid.LoDTensor()