app.py

import gradio as gr
import torch
from transformers import pipeline, AutoModelForSeq2SeqLM, BartTokenizer, AutoModelForCausalLM, AutoTokenizer
from diffusers import StableDiffusionPipeline
from PIL import Image
import io
import os
import zipfile
import traceback

# === Thiết lập thiết bị ===
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {device}")

# === Load models ===
# BART Summarizer
model_name = "lacos03/bart-base-finetuned-xsum"
tokenizer = BartTokenizer.from_pretrained(model_name, use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32
).to(device)
summarizer = pipeline("summarization", model=model, tokenizer=tokenizer, device=0 if device=="cuda" else -1)

# Promptist
promptist_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Promptist",
    torch_dtype=torch.float16 if device == "cuda" else torch.float32
).to(device)
promptist_tokenizer = AutoTokenizer.from_pretrained("gpt2")
promptist_tokenizer.pad_token = promptist_tokenizer.eos_token
promptist_tokenizer.padding_side = "left"

# Stable Diffusion + LoRA
sd_model_id = "runwayml/stable-diffusion-v1-5"
image_generator = StableDiffusionPipeline.from_pretrained(
    sd_model_id,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    use_safetensors=True
).to(device)
lora_weights = "lacos03/std-1.5-lora-midjourney-1.0"
image_generator.load_lora_weights(lora_weights)

# Cấu hình beam search cho BART
NUM_BEAMS = 10
NO_REPEAT_NGRAM_SIZE = 3
LENGTH_PENALTY = 1.0
MIN_NEW_TOKENS = 10
MAX_NEW_TOKENS = 62
# === Hàm xử lý ===
def summarize_article(article_text):
    """Tóm tắt bài viết và tạo prompt refinement"""
    if not article_text.strip():
        return "[Empty]", "[Empty]"
    summary = summarizer(article_text, num_beams=NUM_BEAMS,
        no_repeat_ngram_size=NO_REPEAT_NGRAM_SIZE,
        length_penalty=LENGTH_PENALTY,
        min_new_tokens=MIN_NEW_TOKENS,
        max_new_tokens=MAX_NEW_TOKENS, do_sample=False)[0]["summary_text"]
    title = summary.split(".")[0] + "."
    # Prompt refinement
    input_ids = promptist_tokenizer(title.strip() + " Rephrase:", return_tensors="pt").input_ids.to(device)
    eos_id = promptist_tokenizer.eos_token_id
    outputs = promptist_model.generate(
        input_ids,
        do_sample=False,
        max_new_tokens=75,
        num_beams=8,
        num_return_sequences=1,
        eos_token_id=eos_id,
        pad_token_id=eos_id,
        length_penalty=-1.0
    )
    output_texts = promptist_tokenizer.batch_decode(outputs, skip_special_tokens=True)
    prompt = output_texts[0].replace(title + " Rephrase:", "").strip()
    return title, prompt

def generate_images(prompt, style, num_images=4):
    """Sinh nhiều ảnh"""
    styled_prompt = f"{prompt}, {style.lower()} style"
    results = image_generator(
        styled_prompt,
        num_inference_steps=50,
        guidance_scale=7.5,
        num_images_per_prompt=num_images
    ).images
    return results

def save_selected_images(selected_idx, all_images):
    """Lưu ảnh đã chọn và nén thành ZIP"""
    if not selected_idx:
        return None
    temp_dir = "./temp_selected"
    os.makedirs(temp_dir, exist_ok=True)
    zip_path = os.path.join(temp_dir, "selected_images.zip")
    with zipfile.ZipFile(zip_path, 'w') as zipf:
        for idx in selected_idx:
            img = all_images[int(idx)]
            img_path = os.path.join(temp_dir, f"image_{idx}.png")
            img.save(img_path, format="PNG")
            zipf.write(img_path, f"image_{idx}.png")
    return zip_path

# === UI Gradio ===
def create_app():
    with gr.Blocks() as demo:
        gr.Markdown("## 📰 Article → 🖼️ Multiple Image Generator with Selection")

        # Bước 1: Nhập bài viết và sinh tiêu đề + prompt
        with gr.Row():
            article_input = gr.Textbox(label="📄 Bài viết", lines=10)
            style_dropdown = gr.Dropdown(
                choices=["Art", "Anime", "Watercolor", "Cyberpunk"],
                label="🎨 Phong cách ảnh", value="Art"
            )
        num_images_slider = gr.Slider(1, 8, value=4, step=1, label="🔢 Số lượng ảnh")
        btn_summary = gr.Button("📌 Sinh Tiêu đề & Prompt")

        title_output = gr.Textbox(label="Tiêu đề")
        prompt_output = gr.Textbox(label="Prompt sinh ảnh")

        # Bước 2: Sinh ảnh từ prompt đã refine
        btn_generate_images = gr.Button("🎨 Sinh ảnh từ Prompt")
        gallery = gr.Gallery(label="🖼️ Ảnh minh họa", columns=2, height=600)
        selected_indices = gr.CheckboxGroup(choices=[], label="Chọn ảnh để tải về")

        # Bước 3: Tải ảnh đã chọn
        btn_download = gr.Button("📥 Tải ảnh đã chọn")
        download_file = gr.File(label="File ZIP tải về")

        # Logic
        btn_summary.click(
            fn=summarize_article,
            inputs=[article_input],
            outputs=[title_output, prompt_output]
        )

        def update_gallery(prompt, style, num_images):
            images = generate_images(prompt, style, num_images)
            choices = [str(i) for i in range(len(images))]
            return images, gr.update(choices=choices, value=[]), images  # images lưu tạm trong state

        image_state = gr.State([])

        btn_generate_images.click(
            fn=update_gallery,
            inputs=[prompt_output, style_dropdown, num_images_slider],
            outputs=[gallery, selected_indices, image_state]
        )

        btn_download.click(
            fn=save_selected_images,
            inputs=[selected_indices, image_state],
            outputs=[download_file]
        )

    return demo

# === Chạy app ===
if __name__ == "__main__":
    app = create_app()
    app.launch(debug=True, share=True)