diff --git a/examples/text_to_image/train_text_to_image_lora.py b/examples/text_to_image/train_text_to_image_lora.py
index c8efbddd0b44..afbf80404872 100644
--- a/examples/text_to_image/train_text_to_image_lora.py
+++ b/examples/text_to_image/train_text_to_image_lora.py
@@ -844,10 +844,11 @@ def collate_fn(examples):
                 if args.seed is not None:
                     generator = generator.manual_seed(args.seed)
                 images = []
-                for _ in range(args.num_validation_images):
-                    images.append(
-                        pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
-                    )
+                with torch.cuda.amp.autocast():
+                    for _ in range(args.num_validation_images):
+                        images.append(
+                            pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
+                        )
 
                 for tracker in accelerator.trackers:
                     if tracker.name == "tensorboard":
@@ -913,8 +914,11 @@ def collate_fn(examples):
             if args.seed is not None:
                 generator = generator.manual_seed(args.seed)
             images = []
-            for _ in range(args.num_validation_images):
-                images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
+            with torch.cuda.amp.autocast():
+                for _ in range(args.num_validation_images):
+                    images.append(
+                        pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
+                    )
 
             for tracker in accelerator.trackers:
                 if len(images) != 0: