app.py

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

# ===== CONFIGURATION =====
MODEL_NAME = "HiDream-ai/HiDream-I1-Full"
WATERMARK_TEXT = "SelamGPT"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
TORCH_DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32

# ===== MODEL LOADING =====
# Global variable for model caching (alternative to @gr.Cache)
pipe = None

def load_model():
    global pipe
    if pipe is None:
        pipe = DiffusionPipeline.from_pretrained(
            MODEL_NAME,
            torch_dtype=TORCH_DTYPE
        ).to(DEVICE)
        
        # Optimizations
        if DEVICE == "cuda":
            try:
                pipe.enable_xformers_memory_efficient_attention()
            except:
                print("Xformers not available, using default attention")
            pipe.enable_attention_slicing()
    
    return pipe

# ===== WATERMARK FUNCTION =====
def add_watermark(image):
    """Add watermark with optimized PNG output"""
    try:
        draw = ImageDraw.Draw(image)
        
        font_size = 24
        try:
            font = ImageFont.truetype("Roboto-Bold.ttf", font_size)
        except:
            font = ImageFont.load_default(font_size)
        
        text_width = draw.textlength(WATERMARK_TEXT, font=font)
        x = image.width - text_width - 10
        y = image.height - 34
        
        draw.text((x+1, y+1), WATERMARK_TEXT, font=font, fill=(0, 0, 0, 128))
        draw.text((x, y), WATERMARK_TEXT, font=font, fill=(255, 255, 255))
        
        # Convert to optimized PNG
        img_byte_arr = io.BytesIO()
        image.save(img_byte_arr, format='PNG', optimize=True, quality=85)
        img_byte_arr.seek(0)
        return Image.open(img_byte_arr)
    except Exception as e:
        print(f"Watermark error: {str(e)}")
        return image

# ===== IMAGE GENERATION =====
def generate_image(prompt):
    if not prompt.strip():
        return None, "⚠️ Please enter a prompt"
    
    try:
        model = load_model()
        image = model(
            prompt,
            num_inference_steps=30,
            guidance_scale=7.5
        ).images[0]
        
        return add_watermark(image), "✔️ Generation successful"
    
    except torch.cuda.OutOfMemoryError:
        return None, "⚠️ Out of memory! Try a simpler prompt"
    except Exception as e:
        return None, f"⚠️ Error: {str(e)[:200]}"

# ===== GRADIO INTERFACE =====
# ... (keep your existing interface code exactly as is) ...