Create app.py

app.py

@@ -0,0 +1,120 @@
+import os
+import torch
+import gradio as gr
+from diffusers import SanaPipeline
+
+# Check if GPU is available, fallback safely to CPU
+device = "cuda" if torch.cuda.is_available() else "cpu"
+current_model_id = ""
+pipe = None
+
+def load_sana_model(model_name):
+    global pipe, current_model_id
+    
+    # Map selection names to exact Hugging Face repositories
+    model_map = {
+        "Sana-1.6B (High Quality)": "Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers",
+        "Sana-0.6B (Blazing Fast)": "Efficient-Large-Model/Sana_600M_1024px_diffusers"
+    }
+    
+    selected_id = model_map[model_name]
+    
+    # Avoid reloading if the exact model is already cached in memory
+    if current_model_id == selected_id and pipe is not None:
+        return f"Using already loaded {model_name}"
+        
+    print(f"Loading {selected_id} onto {device}...")
+    
+    try:
+        # Load the pipeline wrapper in bfloat16 to optimize VRAM footprint
+        pipe = SanaPipeline.from_pretrained(
+            selected_id, 
+            torch_dtype=torch.bfloat16 if device == "cuda" else torch.float32
+        )
+        pipe.to(device)
+        
+        # Cast internal sub-networks safely to manage performance bounds
+        if device == "cuda":
+            pipe.vae.to(torch.bfloat16)
+            pipe.text_encoder.to(torch.bfloat16)
+            
+        current_model_id = selected_id
+        return f"Successfully loaded {model_name}!"
+    except Exception as e:
+        return f"Error loading model: {str(e)}"
+
+# Default model initialization
+load_sana_model("Sana-1.6B (High Quality)")
+
+def generate_image(prompt, model_name, steps, guidance_scale, seed):
+    global pipe
+    if pipe is None:
+        load_sana_model(model_name)
+        
+    # Check model compatibility on the fly if the user toggled weights
+    load_sana_model(model_name)
+    
+    # Configure exact deterministic randomness seed
+    if seed == -1:
+        generator = None
+    else:
+        generator = torch.Generator(device=device).manual_seed(int(seed))
+        
+    try:
+        # Sana's specialized autoencoder targets sharp 1024x1024 resolutions natively
+        output = pipe(
+            prompt=prompt,
+            height=1024,
+            width=1024,
+            guidance_scale=float(guidance_scale),
+            num_inference_steps=int(steps),
+            generator=generator
+        )
+        return output.images[0], "Generation completed successfully!"
+    except Exception as e:
+        return None, f"An error occurred: {str(e)}"
+
+# Build custom minimalist theme interface
+css = """
+footer {visibility: hidden !important;}
+#generate_btn {background: linear-gradient(90deg, #10b981, #059669) !important; color: white !important;}
+"""
+
+with gr.Blocks(theme=gr.themes.Base(), css=css) as demo:
+    gr.Markdown("# 🚀 Sana Linear-Attention Image Engine")
+    gr.Markdown("Deeply optimized text-to-image generation running instantly via pure diffusers implementation.")
+    
+    with gr.Row():
+        with gr.Column(scale=4):
+            prompt_input = gr.Textbox(
+                label="Your Prompt String", 
+                placeholder="A cyberpunk character portrait, glowing neon line accents, minimalist black aesthetic...", 
+                lines=3
+            )
+            model_selector = gr.Radio(
+                choices=["Sana-1.6B (High Quality)", "Sana-0.6B (Blazing Fast)"], 
+                value="Sana-1.6B (High Quality)", 
+                label="Active Core Architecture"
+            )
+            
+            with gr.Accordion("Advanced Hyperparameters", open=False):
+                inference_steps = gr.Slider(minimum=4, maximum=50, value=20, step=1, label="Inference Loop Steps")
+                cfg_scale = gr.Slider(minimum=1.0, maximum=10.0, value=4.5, step=0.1, label="Guidance Scale (CFG)")
+                random_seed = gr.Number(value=-1, label="Manual Seed Value (-1 for Random)")
+                
+            generate_button = gr.Button("Synthesize Frame Assets", variant="primary", elem_id="generate_btn")
+            status_output = gr.Textbox(label="Backend System Logging", interactive=False)
+            
+        with gr.Column(scale=5):
+            image_output = gr.Image(label="Render Output Pipeline", type="pil", interactive=False)
+
+    # Bind active logic flows
+    generate_button.click(
+        fn=generate_image,
+        inputs=[prompt_input, model_selector, inference_steps, cfg_scale, random_seed],
+        outputs=[image_output, status_output]
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch()
+