app.py

import os
import torch
import gradio as gr
from diffusers import DiffusionPipeline
from PIL import Image, ImageDraw, ImageFont

# ===== FREE-TIER CONFIG =====
WATERMARK_TEXT = "SelamGPT"
MODEL_NAME = "DeepFloyd/IF-II-L-v1.0"
CACHE_DIR = "model_cache"  # For free tier storage limits

# ===== LIGHTWEIGHT MODEL LOAD =====
pipe = None  # Lazy load to avoid cold start timeouts

def load_model():
    global pipe
    if pipe is None:
        pipe = DiffusionPipeline.from_pretrained(
            MODEL_NAME,
            torch_dtype=torch.float16,  # 50% VRAM reduction
            variant="fp16",
            cache_dir=CACHE_DIR
        )
        pipe.enable_model_cpu_offload()  # Critical for free-tier VRAM

# ===== OPTIMIZED WATERMARK =====
def add_watermark(image):
    try:
        draw = ImageDraw.Draw(image)
        font = ImageFont.load_default(20)  # No external font needed
        text_width = draw.textlength(WATERMARK_TEXT, font=font)
        draw.text(
            (image.width - text_width - 15, image.height - 30),
            WATERMARK_TEXT,
            font=font,
            fill=(255, 255, 255)
        )
        return image
    except Exception:
        return image

# ===== FREE-TIER GENERATION =====
def generate_image(prompt):
    if not prompt.strip():
        return None, "⚠️ Please enter a prompt"
    
    try:
        load_model()  # Lazy load only when needed
        
        # Free-tier optimized settings
        result = pipe(
            prompt=prompt,
            output_type="pil",
            generator=torch.Generator().manual_seed(42),  # Consistent results
            num_inference_steps=30,  # Reduced from default 50
            guidance_scale=7.0  # Balanced creativity/quality
        )
        
        return add_watermark(result.images[0]), "✔️ Generated (Free Tier)"
    except torch.cuda.OutOfMemoryError:
        return None, "⚠️ Out of VRAM - Try simpler prompt"
    except Exception as e:
        return None, f"⚠️ Error: {str(e)[:100]}"

# ===== GRADIO UI =====
with gr.Blocks(title="SelamGPT Pro") as demo:
    gr.Markdown("""
    # 🎨 SelamGPT (DeepFloyd IF-II-L)
    *Optimized for Free Tier - 64px Base Resolution*
    """)
    
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(
                label="Describe your image",
                placeholder="A traditional Ethiopian market...",
                lines=3
            )
            generate_btn = gr.Button("Generate", variant="primary")
            
            gr.Examples(
                examples=[
                    ["Habesha cultural dress with intricate patterns, studio lighting"],
                    ["Lalibela rock-hewn churches at golden hour"],
                    ["Addis Ababa futuristic skyline, cyberpunk style"]
                ],
                inputs=prompt_input
            )
            
        with gr.Column():
            output_image = gr.Image(
                label="Generated Image",
                type="pil",
                format="webp",  # Lightweight format
                height=400
            )
            status_output = gr.Textbox(
                label="Status",
                interactive=False
            )
    
    generate_btn.click(
        fn=generate_image,
        inputs=prompt_input,
        outputs=[output_image, status_output]
    )

if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        enable_queue=False  # Critical for free tier
    )