Mistral-7B-Instruct-v0.2.py

# import gradio as gr
# from transformers import (
#     Blip2Processor,
#     Blip2ForConditionalGeneration,
#     AutoTokenizer,
#     AutoModelForCausalLM,
# )
# from PIL import Image
# import torch

# # Set device
# device = "cuda" if torch.cuda.is_available() else "cpu"

# # Load image captioning model (BLIP-2)
# processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
# blip_model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xl").to(device)

# # Load text generation model (LLM)
# llm_tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
# llm_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2").to(device)

# # Step 1: Generate image caption
# def extract_caption(image):
#     inputs = processor(images=image, return_tensors="pt").to(device)
#     outputs = blip_model.generate(**inputs, max_new_tokens=50)
#     caption = processor.tokenizer.decode(outputs[0], skip_special_tokens=True)
#     return caption

# # Step 2: Build fairytale prompt
# def build_prompt(caption):
#     return (
#         f"Based on the image description: \"{caption}\", write a children's fairytale.\n"
#         "The story must:\n"
#         "- Start with 'Once upon a time'\n"
#         "- Be at least 10 full sentences long\n"
#         "- Include named characters, a clear setting, emotions, a challenge, and a resolution\n"
#         "- Avoid mentions of babies or unrelated royalty unless relevant\n"
#         "Here is the story:\nOnce upon a time"
#     )

# # Step 3: Generate story
# def generate_fairytale(image):
#     caption = extract_caption(image)
#     prompt = build_prompt(caption)
#     inputs = llm_tokenizer(prompt, return_tensors="pt").to(device)

#     output = llm_model.generate(
#         **inputs,
#         max_new_tokens=500,
#         do_sample=True,
#         temperature=0.9,
#         top_p=0.95,
#         pad_token_id=llm_tokenizer.eos_token_id
#     )

#     result = llm_tokenizer.decode(output[0], skip_special_tokens=True)

#     # Trim to only the story
#     if "Once upon a time" in result:
#         return "Once upon a time" + result.split("Once upon a time", 1)[-1].strip()
#     else:
#         return f"⚠️ Failed to generate story.\n\n[Prompt]\n{prompt}\n\n[Output]\n{result}"

# # Gradio interface
# with gr.Blocks() as demo:
#     gr.Markdown("## 📖 AI Fairytale Generator\nUpload an image and get a magical story!")

#     with gr.Row():
#         image_input = gr.Image(type="pil", label="Upload an image")

#     with gr.Row():
#         generate_button = gr.Button("✨ Generate Fairytale")

#     with gr.Row():
#         output_text = gr.Textbox(label="Generated Story", lines=20)

#     generate_button.click(fn=generate_fairytale, inputs=[image_input], outputs=[output_text])

# if __name__ == "__main__":
#     demo.launch(share=True)

import gradio as gr
from transformers import (
    Blip2Processor,
    Blip2ForConditionalGeneration,
    AutoTokenizer,
    AutoModelForCausalLM,
)
from PIL import Image
import torch

# 디바이스 설정 (Mistral은 CPU로 강제 설정)
device = "cpu"

# 1️⃣ 이미지 설명 생성 모델 로드 (BLIP-2)
processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
blip_model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xl").to(device)

# 2️⃣ 동화 생성 모델 로드 (Mistral-7B, CPU)
llm_tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
llm_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", device_map="cpu")

# 3️⃣ 이미지 → 설명
def extract_caption(image):
    inputs = processor(images=image, return_tensors="pt").to(device)
    outputs = blip_model.generate(**inputs, max_new_tokens=50)
    caption = processor.tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
    return caption

# 4️⃣ 프롬프트 구성
def build_prompt(caption):
    return (
        f"Based on the image description: \"{caption}\", write a children's fairytale.\n"
        "The story must:\n"
        "- Start with 'Once upon a time'\n"
        "- Be at least 10 full sentences long\n"
        "- Include named characters, a clear setting, emotions, a challenge, and a resolution\n"
        "- Avoid unrelated royalty or babies unless relevant\n"
        "Here is the story:\nOnce upon a time"
    )

# 5️⃣ 전체 동화 생성
def generate_fairytale(image):
    caption = extract_caption(image)
    prompt = build_prompt(caption)

    inputs = llm_tokenizer(prompt, return_tensors="pt").to(device)

    output = llm_model.generate(
        **inputs,
        max_new_tokens=500,
        do_sample=True,
        temperature=0.9,
        top_p=0.95,
        pad_token_id=llm_tokenizer.eos_token_id
    )

    result = llm_tokenizer.decode(output[0], skip_special_tokens=True)

    # "Once upon a time" 이후만 추출
    if "Once upon a time" in result:
        return caption, "Once upon a time" + result.split("Once upon a time", 1)[-1].strip()
    else:
        return caption, f"⚠️ Story generation failed.\n\n[Prompt]\n{prompt}\n\n[Output]\n{result}"

# 6️⃣ Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## 🧚 AI Fairytale Generator (Mistral CPU ver.)")
    gr.Markdown("Upload an image and receive a magical children's fairytale based on it ✨")

    image_input = gr.Image(type="pil", label="🖼️ Upload an Image")
    generate_button = gr.Button("✨ Generate Fairytale")

    caption_output = gr.Textbox(label="📌 Image Description", lines=2)
    story_output = gr.Textbox(label="📖 Generated Fairytale", lines=20)

    generate_button.click(fn=generate_fairytale, inputs=[image_input], outputs=[caption_output, story_output])

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