Speed Increased by loading pipeline

I increased speed of this demo by loading pipeline once instead of each generation.

app.py

@@ -33,11 +33,10 @@ ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
-def load_pipeline(pipeline_type):
-    if pipeline_type == "text2img":
-        return StableDiffusion3Pipeline.from_pretrained(model_path, torch_dtype=torch.float16)
-    elif pipeline_type == "img2img":
-        return StableDiffusion3Img2ImgPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
+pipe_t2i = StableDiffusion3Pipeline.from_pretrained(model_path, torch_dtype=torch.float16)
+pipe_t2i.to(device)
+pipe_i2i = StableDiffusion3Img2ImgPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
+pipe_i2i.tp(device)
 
 def save_image(img):
     unique_name = str(uuid.uuid4()) + ".png"
@@ -66,15 +65,13 @@ def generate(
     use_resolution_binning: bool = True,
     progress=gr.Progress(track_tqdm=True),
 ):
-    pipe = load_pipeline("text2img")
-    pipe.to(device)
     seed = int(randomize_seed_fn(seed, randomize_seed))
     generator = torch.Generator().manual_seed(seed)
     
     if not use_negative_prompt:
         negative_prompt = None # type: ignore
     
-    output = pipe(
+    output = pipe_t2i(
         prompt=prompt,
         negative_prompt=negative_prompt,
         width=width,
@@ -104,8 +101,6 @@ def img2img_generate(
     use_resolution_binning: bool = True,
     progress=gr.Progress(track_tqdm=True),
 ):
-    pipe = load_pipeline("img2img")
-    pipe.to(device)
     seed = int(randomize_seed_fn(seed, randomize_seed))
     generator = torch.Generator().manual_seed(seed)
     
@@ -114,7 +109,7 @@ def img2img_generate(
     
     init_image = init_image.resize((768, 768))
     
-    output = pipe(
+    output = pipe_i2i(
         prompt=prompt,
         image=init_image,
         negative_prompt=negative_prompt,