diff --git a/onnxruntime/python/tools/transformers/models/stable_diffusion/pipeline_txt2img.py b/onnxruntime/python/tools/transformers/models/stable_diffusion/pipeline_txt2img.py
index 444b6d9a8ca14..b9759b44e7635 100644
--- a/onnxruntime/python/tools/transformers/models/stable_diffusion/pipeline_txt2img.py
+++ b/onnxruntime/python/tools/transformers/models/stable_diffusion/pipeline_txt2img.py
@@ -79,7 +79,7 @@ def _infer(
             latents = self.denoise_latent(latents, text_embeddings, guidance=guidance)
 
             # VAE decode latent
-            images = self.decode_latent(latents)
+            images = self.decode_latent(latents / self.vae_scaling_factor)
 
             torch.cuda.synchronize()
             e2e_toc = time.perf_counter()