app.py

import spaces
import gradio as gr
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
from PIL import Image
import torch
import os, time
import logging


logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Load the model and processor
# model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
#     "daniel3303/QwenStoryteller",
#     torch_dtype=torch.float16,
#     device_map="auto"
# )
# processor = AutoProcessor.from_pretrained("daniel3303/QwenStoryteller")
# Load the model and processor
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "daniel3303/QwenStoryteller2",
    torch_dtype=torch.float16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("daniel3303/QwenStoryteller2")

@spaces.GPU(duration=180)
@torch.no_grad()
def generate_story(file_paths, progress=gr.Progress(track_tqdm=True)):
    # Load images from the file paths
    images = [Image.open(file_path) for file_path in file_paths]
    
    image_content = []
    for img in images[:10]:  # Limit to 6 images
        image_content.append({
            "type": "image",
            "image": img,
        })

    image_content.append({"type": "text", "text": "Generate a story based on these images."})

    system = """You are an AI storyteller that can analyze sequences of images and create creative narratives. 
    First think step-by-step to analyze characters, objects, settings, and narrative structure. 
    Then create a grounded story that maintains consistent character identity and object references across frames. 
    Use <think></think> tags to show your reasoning process before writing the final story.
    """
    messages = [
        {
            "role": "system",
            "content": system
        },
        {
            "role": "user",
            "content": image_content,
        }
    ]

    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    image_inputs, video_inputs = process_vision_info(messages)
    inputs = processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt"
    )
    inputs = inputs.to(model.device)

    generated_ids = model.generate(
        **inputs,
        max_new_tokens=4096,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )

    generated_ids_trimmed = [
        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
    ]
    story = processor.batch_decode(
        generated_ids_trimmed,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False
    )[0]

    return story


with gr.Blocks(fill_height=True) as demo:
    gr.Markdown("# Qwen Storyteller \n## Upload up to 10 images to generate a creative story.")
    
    with gr.Row():
        with gr.Column(scale=1):
            upload_button = gr.UploadButton("Upload up to 10 images", file_types=["image"], file_count="multiple")
            output_file = gr.File(label="Uploaded Files")
            gen_button = gr.Button("Generate", variant="primary")
            gr.Markdown(
                """
                \n\n
                ```
                @misc{oliveira2025storyreasoningdatasetusingchainofthought,
                      title={StoryReasoning Dataset: Using Chain-of-Thought for Scene Understanding and Grounded Story Generation}, 
                      author={Daniel A. P. Oliveira and David Martins de Matos},
                      year={2025},
                      eprint={2505.10292},
                      archivePrefix={arXiv},
                      primaryClass={cs.CV},
                      url={https://arxiv.org/abs/2505.10292}, 
                }
                ```
                """
            )
        with gr.Column(scale=2):
            outputs = gr.Markdown(label="Generated Story", show_copy_button=True)

    with gr.Row():
        gr.Markdown(
            """
            ### Key Features
            * Cross-Frame Consistency: Maintains consistent character and object identity across multiple frames through visual similarity and face recognition techniques
            * Structured Reasoning: Employs chain-of-thought reasoning to analyze scenes with explicit modeling of characters, objects, settings, and narrative structure
            * Grounded Storytelling: Uses specialized XML tags to link narrative elements directly to visual entities
            * Reduced Hallucinations: Achieves 12.3% fewer hallucinations compared to the non-fine-tuned base model
            
            Model trained by daniel3303, [repository here.](https://huggingface.co/daniel3303/QwenStoryteller2)
            """    
        )

        
    upload_button.upload(lambda files: [f.name for f in files], upload_button, output_file)

    gen_button.click(generate_story, upload_button, outputs)

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