Deploy Gradio app with multiple files

app.py

@@ -0,0 +1,31 @@
+import gradio as gr
+from models import load_model, generate_image
+
+# Load the model on startup
+pipe = load_model()
+
+def text_to_image(prompt):
+    return generate_image(pipe, prompt)
+
+with gr.Blocks(title="AI Text-to-Image Generator") as demo:
+    gr.Markdown("# AI Text-to-Image Generator\nBuilt with [anycoder](https://huggingface.co/spaces/akhaliq/anycoder)")
+    gr.Markdown("Generate images from text prompts using FLUX.1-dev model.")
+    
+    prompt_input = gr.Textbox(
+        label="Enter your prompt",
+        placeholder="A beautiful landscape with mountains and a lake...",
+        lines=3
+    )
+    
+    generate_btn = gr.Button("Generate Image")
+    
+    output_image = gr.Image(label="Generated Image")
+    
+    generate_btn.click(
+        fn=text_to_image,
+        inputs=[prompt_input],
+        outputs=[output_image]
+    )
+
+if __name__ == "__main__":
+    demo.launch()

config.py

@@ -0,0 +1,7 @@
+# Configuration constants
+MODEL_ID = 'black-forest-labs/FLUX.1-dev'
+DEFAULT_NUM_STEPS = 20
+DEFAULT_GUIDANCE_SCALE = 3.5
+DEFAULT_HEIGHT = 512
+DEFAULT_WIDTH = 512
+MAX_PROMPT_LENGTH = 500

models.py

@@ -0,0 +1,49 @@
+import spaces
+import torch
+from diffusers import DiffusionPipeline
+
+MODEL_ID = 'black-forest-labs/FLUX.1-dev'
+
+# Compile the model ahead-of-time for optimal performance
+@spaces.GPU(duration=1500)  # Maximum duration for compilation
+def compile_transformer():
+    pipe = DiffusionPipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
+    pipe.to('cuda')
+    
+    # Capture example inputs for AoT compilation
+    with spaces.aoti_capture(pipe.transformer) as call:
+        pipe("arbitrary example prompt")
+    
+    # Export the model
+    exported = torch.export.export(
+        pipe.transformer,
+        args=call.args,
+        kwargs=call.kwargs,
+    )
+    
+    # Compile the exported model
+    return spaces.aoti_compile(exported)
+
+# Load and compile the model
+def load_model():
+    pipe = DiffusionPipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
+    pipe.to('cuda')
+    
+    # Apply compiled transformer
+    compiled_transformer = compile_transformer()
+    spaces.aoti_apply(compiled_transformer, pipe.transformer)
+    
+    return pipe
+
+@spaces.GPU  # GPU inference for fast generation
+def generate_image(pipe, prompt):
+    # Generate image with optimized settings
+    image = pipe(
+        prompt,
+        num_inference_steps=20,  # Fewer steps for speed
+        guidance_scale=3.5,
+        height=512,
+        width=512
+    ).images[0]
+    
+    return image

requirements.txt

@@ -0,0 +1,10 @@
+gradio
+git+https://github.com/huggingface/diffusers
+torch
+accelerate
+transformers
+spaces
+safetensors
+numpy
+Pillow
+huggingface_hub

utils.py

@@ -0,0 +1,8 @@
+# Utility functions for the text-to-image generator
+def validate_prompt(prompt):
+    """Basic prompt validation"""
+    if not prompt or len(prompt.strip()) == 0:
+        return False, "Prompt cannot be empty"
+    if len(prompt) > 500:
+        return False, "Prompt is too long (max 500 characters)"
+    return True, ""