app.py

from transformers import Blip2Processor, Blip2ForConditionalGeneration, AutoModelForCausalLM, AutoTokenizer
import torch
from PIL import Image, ImageDraw, ImageFont
import gradio as gr
import torch
import os

# Check CUDA availability
if not torch.cuda.is_available():
    os.environ["BITSANDBYTES_NOWELCOME"] = "1"
    os.environ["CUDA_VISIBLE_DEVICES"] = ""
    os.environ["LIBRARY_PATH"] = "/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

# Initialize environment
os.makedirs("./offload", exist_ok=True)
HF_TOKEN = os.environ.get("HF_TOKEN")

# Memory optimization
torch.cuda.empty_cache()
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True

# Load BLIP-2
blip_processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
blip_model = Blip2ForConditionalGeneration.from_pretrained(
    "Salesforce/blip2-opt-2.7b",
    torch_dtype=torch.float16,
    device_map="auto"
).eval()

# Load Phi-3
phi_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    trust_remote_code=True,
    device_map="auto",
    torch_dtype=torch.float16,
    load_in_4bit=torch.cuda.is_available(),  # Only use 4bit if CUDA available
    token=HF_TOKEN
).eval()
phi_tokenizer = AutoTokenizer.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    token=HF_TOKEN
)

def analyze_image(image):
    inputs = blip_processor(image, return_tensors="pt").to(blip_model.device)
    generated_ids = blip_model.generate(**inputs, max_length=50)
    return blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

def generate_meme_caption(image_desc, user_prompt):
    messages = [
        {"role": "system", "content": "You are a meme expert. Create funny captions in format: TOP TEXT | BOTTOM TEXT"},
        {"role": "user", "content": f"Image context: {image_desc}\nUser input: {user_prompt}\nGenerate 3 meme captions (max 10 words each):"}
    ]
    
    inputs = phi_tokenizer.apply_chat_template(
        messages,
        return_tensors="pt",
        add_generation_prompt=True
    ).to(phi_model.device)
    
    outputs = phi_model.generate(
        inputs,
        max_new_tokens=200,
        temperature=0.7,
        do_sample=True
    )
    return phi_tokenizer.decode(outputs[0], skip_special_tokens=True)

def create_meme(image, top_text, bottom_text):
    img = image.copy()
    draw = ImageDraw.Draw(img)
    
    # Use available font (works in Colab/Spaces)
    try:
        font = ImageFont.truetype("arial.ttf", size=min(img.size)//12)
    except:
        font = ImageFont.load_default()
    
    # Top text
    draw.text(
        (img.width/2, 10), 
        top_text, 
        font=font, 
        fill="white", 
        anchor="mt",
        stroke_width=2, 
        stroke_fill="black"
    )
    
    # Bottom text
    draw.text(
        (img.width/2, img.height-10), 
        bottom_text, 
        font=font, 
        fill="white", 
        anchor="mb",
        stroke_width=2, 
        stroke_fill="black"
    )
    return img

def process_meme(image, user_prompt):
    image_desc = analyze_image(image)
    raw_output = generate_meme_caption(image_desc, user_prompt)
    
    captions = []
    for line in raw_output.split("\n"):
        if "|" in line:
            parts = line.split("|", 1)
            if len(parts) == 2:
                captions.append((parts[0].strip(), parts[1].strip()))
    
    memes = [create_meme(image, top, bottom) for top, bottom in captions[:3]]
    return memes

with gr.Blocks(title="AI Meme Generator") as demo:
    gr.Markdown("# 🚀 AI Meme Generator")
    with gr.Row():
        image_input = gr.Image(type="pil", label="Upload Image")
        text_input = gr.Textbox(label="Meme Theme/Prompt")
    submit_btn = gr.Button("Generate Memes!")
    gallery = gr.Gallery(label="Generated Memes", columns=3)
    
    submit_btn.click(
        fn=process_meme,
        inputs=[image_input, text_input],
        outputs=gallery
    )

if __name__ == "__main__":
    demo.launch()