Refactor app.py to initialize FLUX pipeline and VAE on CUDA, enhancing image generation capabilities

app.py

@@ -1,3 +1,10 @@
+from diffusers import FluxPipeline, AutoencoderKL
+from live_preview_helpers import flux_pipe_call_that_returns_an_iterable_of_images
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to(device)
+good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=torch.bfloat16).to(device)
+pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)
+
 import gradio as gr
 import spaces
 from gradio_litmodel3d import LitModel3D
@@ -15,15 +22,9 @@ from trellis.pipelines import TrellisImageTo3DPipeline
 from trellis.representations import Gaussian, MeshExtractResult
 from trellis.utils import render_utils, postprocessing_utils
 from gradio_client import Client
-from diffusers import FluxPipeline, AutoencoderKL
-from live_preview_helpers import flux_pipe_call_that_returns_an_iterable_of_images
 
 llm_client = Client("Qwen/Qwen2.5-72B-Instruct")
 
-pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
-good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=torch.bfloat16)
-pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)
-
 def generate_t2i_prompt(item_name):
     llm_prompt_template = """You are tasked with creating a concise yet highly detailed description of an item to be used for generating an image in a game development pipeline. The image should show the **entire item** with no parts cropped or hidden. The background should always be plain and monocolor, with no focus on it.  
 
@@ -56,7 +57,6 @@ Focus on the item itself, ensuring it is fully described, and specify a plain, w
 @spaces.GPU(duration=75)
 def generate_item_image(object_t2i_prompt):
     trial_id = ""
-    pipe.cuda()
     for image in pipe.flux_pipe_call_that_returns_an_iterable_of_images(
         prompt=object_t2i_prompt,
         guidance_scale=3.5,