Update app.py

app.py

@@ -1,39 +1,41 @@
 import gradio as gr
 import numpy as np
 import random
-from diffusers import DiffusionPipeline, DDPMScheduler, DPMSolverMultistepScheduler
 import torch
+from diffusers import DiffusionPipeline, EDMEulerScheduler
 
+# 🖥️ Detect device
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
+# 🎯 Model ID and config
+model_repo_id = "stabilityai/sdxl-turbo"
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 1024
 
-# 🧮 Use ODE-based DPM-SolverMultistepScheduler for faster inference
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
+# 🔁 Load model with EDM + VPred scheduler
+pipe = DiffusionPipeline.from_pretrained(
+    model_repo_id,
+    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+    variant="fp16" if torch.cuda.is_available() else None,
+)
 
-# ✅ Replace default scheduler with an ODE-optimized scheduler
-pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+# 🔄 Replace scheduler with EDM + V-prediction
+pipe.scheduler = EDMEulerScheduler.from_config(pipe.scheduler.config)
 
-# 🚀 Optimization Flags (Math-Driven)
-try:
-    pipe.enable_xformers_memory_efficient_attention()
-except Exception as e:
-    print("xFormers not available, falling back to slicing.")
-    pipe.enable_attention_slicing()
+# 🧠 Enable optimizations if GPU
+if device == "cuda":
+    try:
+        pipe.enable_xformers_memory_efficient_attention()
+    except Exception as e:
+        print("⚠️ xFormers not available, using attention slicing.")
+        pipe.enable_attention_slicing()
 
-pipe.enable_model_cpu_offload()
-pipe.enable_attention_slicing()
-pipe.enable_vae_tiling()
+    pipe.enable_model_cpu_offload()
+    pipe.enable_vae_tiling()
 
 pipe = pipe.to(device)
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
-
+# 🚀 Inference function
 def infer(
     prompt,
     negative_prompt,
@@ -48,7 +50,7 @@ def infer(
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
 
-    generator = torch.Generator().manual_seed(seed)
+    generator = torch.Generator(device=device).manual_seed(seed)
 
     image = pipe(
         prompt=prompt,
@@ -62,12 +64,14 @@ def infer(
 
     return image, seed
 
+# 🧪 Prompt examples
 examples = [
     "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
     "An astronaut riding a green horse",
     "A delicious ceviche cheesecake slice",
 ]
 
+# 🎨 UI CSS
 css = """
 #col-container {
     margin: 0 auto;
@@ -75,9 +79,10 @@ css = """
 }
 """
 
+# 🧱 Gradio Interface
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
+        gr.Markdown(" # Text-to-Image Gradio (EDM + VPred)")
 
         with gr.Row():
             prompt = gr.Text(
@@ -133,18 +138,19 @@ with gr.Blocks(css=css) as demo:
                     minimum=0.0,
                     maximum=10.0,
                     step=0.1,
-                    value=0.0,
+                    value=2.5,  # Optimal for SDXL-Turbo
                 )
 
                 num_inference_steps = gr.Slider(
                     label="Number of inference steps",
                     minimum=1,
-                    maximum=50,
+                    maximum=20,
                     step=1,
-                    value=2,
+                    value=4,  # Low default for EDM
                 )
 
         gr.Examples(examples=examples, inputs=[prompt])
+
     gr.on(
         triggers=[run_button.click, prompt.submit],
         fn=infer,