Needle-2B-VL-Highlights

The Needle-2B-VL-Highlights model is a fine-tuned version of Qwen2-VL-2B-Instruct, specifically optimized for image highlights extraction, messy handwriting recognition, Optical Character Recognition (OCR), English language understanding, and math problem solving with LaTeX formatting. This model uses a conversational visual-language interface to effectively handle multi-modal tasks.

Key Enhancements:

• State-of-the-art image comprehension across varying resolutions and aspect ratios:
  Needle-2B-VL-Highlights delivers top-tier performance on benchmarks such as MathVista, DocVQA, RealWorldQA, and MTVQA.

• Image Highlighting Expertise:
  Specially tuned to identify and summarize key visual elements in an image — ideal for creating visual highlights, annotations, and summaries.

• Handwriting OCR Enhanced:
  Recognizes messy and complex handwritten notes with precision, perfect for digitizing real-world documents.

• Video Content Understanding:
  Capable of processing videos longer than 20 minutes for context-aware Q&A, transcription, and highlight extraction.

• Multi-device Integration:
  Can be used as an intelligent agent for mobile phones, robots, and other devices — able to understand visual scenes and execute actions.

• Multilingual OCR Support:
  In addition to English and Chinese, supports OCR for European languages, Japanese, Korean, Arabic, and Vietnamese.

Run with Transformers🤗

%%capture
!pip install -q gradio spaces transformers accelerate
!pip install -q numpy requests torch torchvision
!pip install -q qwen-vl-utils av ipython reportlab
!pip install -q fpdf python-docx pillow huggingface_hub

#Demo
import gradio as gr
import spaces
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, TextIteratorStreamer
from qwen_vl_utils import process_vision_info
import torch
from PIL import Image
import os
import uuid
import io
from threading import Thread
from reportlab.lib.pagesizes import A4
from reportlab.lib.styles import getSampleStyleSheet
from reportlab.lib import colors
from reportlab.platypus import SimpleDocTemplate, Image as RLImage, Paragraph, Spacer
from reportlab.lib.units import inch
from reportlab.pdfbase import pdfmetrics
from reportlab.pdfbase.ttfonts import TTFont
import docx
from docx.enum.text import WD_ALIGN_PARAGRAPH

# Define model options
MODEL_OPTIONS = {
    "Needle-2B-VL-Highlights": "prithivMLmods/Needle-2B-VL-Highlights",
}

# Preload models and processors into CUDA
models = {}
processors = {}
for name, model_id in MODEL_OPTIONS.items():
    print(f"Loading {name}...")
    models[name] = Qwen2VLForConditionalGeneration.from_pretrained(
        model_id,
        trust_remote_code=True,
        torch_dtype=torch.float16
    ).to("cuda").eval()
    processors[name] = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

image_extensions = Image.registered_extensions()

def identify_and_save_blob(blob_path):
    """Identifies if the blob is an image and saves it."""
    try:
        with open(blob_path, 'rb') as file:
            blob_content = file.read()
            try:
                Image.open(io.BytesIO(blob_content)).verify()  # Check if it's a valid image
                extension = ".png"  # Default to PNG for saving
                media_type = "image"
            except (IOError, SyntaxError):
                raise ValueError("Unsupported media type. Please upload a valid image.")

            filename = f"temp_{uuid.uuid4()}_media{extension}"
            with open(filename, "wb") as f:
                f.write(blob_content)

            return filename, media_type

    except FileNotFoundError:
        raise ValueError(f"The file {blob_path} was not found.")
    except Exception as e:
        raise ValueError(f"An error occurred while processing the file: {e}")

@spaces.GPU
def qwen_inference(model_name, media_input, text_input=None):
    """Handles inference for the selected model."""
    model = models[model_name]
    processor = processors[model_name]

    if isinstance(media_input, str):
        media_path = media_input
        if media_path.endswith(tuple([i for i in image_extensions.keys()])):
            media_type = "image"
        else:
            try:
                media_path, media_type = identify_and_save_blob(media_input)
            except Exception as e:
                raise ValueError("Unsupported media type. Please upload a valid image.")

    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": media_type,
                    media_type: media_path
                },
                {"type": "text", "text": text_input},
            ],
        }
    ]

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

    streamer = TextIteratorStreamer(
        processor.tokenizer, skip_prompt=True, skip_special_tokens=True
    )
    generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)

    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    buffer = ""
    for new_text in streamer:
        buffer += new_text
        # Remove <|im_end|> or similar tokens from the output
        buffer = buffer.replace("<|im_end|>", "")
        yield buffer

def format_plain_text(output_text):
    """Formats the output text as plain text without LaTeX delimiters."""
    # Remove LaTeX delimiters and convert to plain text
    plain_text = output_text.replace("\\(", "").replace("\\)", "").replace("\\[", "").replace("\\]", "")
    return plain_text

def generate_document(media_path, output_text, file_format, font_size, line_spacing, alignment, image_size):
    """Generates a document with the input image and plain text output."""
    plain_text = format_plain_text(output_text)
    if file_format == "pdf":
        return generate_pdf(media_path, plain_text, font_size, line_spacing, alignment, image_size)
    elif file_format == "docx":
        return generate_docx(media_path, plain_text, font_size, line_spacing, alignment, image_size)

def generate_pdf(media_path, plain_text, font_size, line_spacing, alignment, image_size):
    """Generates a PDF document."""
    filename = f"output_{uuid.uuid4()}.pdf"
    doc = SimpleDocTemplate(
        filename,
        pagesize=A4,
        rightMargin=inch,
        leftMargin=inch,
        topMargin=inch,
        bottomMargin=inch
    )
    styles = getSampleStyleSheet()
    styles["Normal"].fontSize = int(font_size)
    styles["Normal"].leading = int(font_size) * line_spacing
    styles["Normal"].alignment = {
        "Left": 0,
        "Center": 1,
        "Right": 2,
        "Justified": 4
    }[alignment]

    story = []

    # Add image with size adjustment
    image_sizes = {
        "Small": (200, 200),
        "Medium": (400, 400),
        "Large": (600, 600)
    }
    img = RLImage(media_path, width=image_sizes[image_size][0], height=image_sizes[image_size][1])
    story.append(img)
    story.append(Spacer(1, 12))

    # Add plain text output
    text = Paragraph(plain_text, styles["Normal"])
    story.append(text)

    doc.build(story)
    return filename

def generate_docx(media_path, plain_text, font_size, line_spacing, alignment, image_size):
    """Generates a DOCX document."""
    filename = f"output_{uuid.uuid4()}.docx"
    doc = docx.Document()

    # Add image with size adjustment
    image_sizes = {
        "Small": docx.shared.Inches(2),
        "Medium": docx.shared.Inches(4),
        "Large": docx.shared.Inches(6)
    }
    doc.add_picture(media_path, width=image_sizes[image_size])
    doc.add_paragraph()

    # Add plain text output
    paragraph = doc.add_paragraph()
    paragraph.paragraph_format.line_spacing = line_spacing
    paragraph.paragraph_format.alignment = {
        "Left": WD_ALIGN_PARAGRAPH.LEFT,
        "Center": WD_ALIGN_PARAGRAPH.CENTER,
        "Right": WD_ALIGN_PARAGRAPH.RIGHT,
        "Justified": WD_ALIGN_PARAGRAPH.JUSTIFY
    }[alignment]
    run = paragraph.add_run(plain_text)
    run.font.size = docx.shared.Pt(int(font_size))

    doc.save(filename)
    return filename

# CSS for output styling
css = """
  #output {
    height: 500px;
    overflow: auto;
    border: 1px solid #ccc;
  }
.submit-btn {
    background-color: #cf3434  !important;
    color: white !important;
}
.submit-btn:hover {
    background-color: #ff2323 !important;
}
.download-btn {
    background-color: #35a6d6 !important;
    color: white !important;
}
.download-btn:hover {
    background-color: #22bcff !important;
}
"""

# Gradio app setup
with gr.Blocks(css=css) as demo:
    gr.Markdown("# Qwen2VL Models: Vision and Language Processing")

    with gr.Tab(label="Image Input"):

        with gr.Row():
            with gr.Column():
                model_choice = gr.Dropdown(
                    label="Model Selection",
                    choices=list(MODEL_OPTIONS.keys()),
                    value="Needle-2B-VL-Highlights"
                )
                input_media = gr.File(
                    label="Upload Image", type="filepath"
                )
                text_input = gr.Textbox(label="Question", placeholder="Ask a question about the image...")
                submit_btn = gr.Button(value="Submit", elem_classes="submit-btn")

            with gr.Column():
                output_text = gr.Textbox(label="Output Text", lines=10)
                plain_text_output = gr.Textbox(label="Standardized Plain Text", lines=10)

        submit_btn.click(
            qwen_inference, [model_choice, input_media, text_input], [output_text]
        ).then(
            lambda output_text: format_plain_text(output_text), [output_text], [plain_text_output]
        )

        # Add examples directly usable by clicking
        with gr.Row():
            with gr.Column():
                line_spacing = gr.Dropdown(
                    choices=[0.5, 1.0, 1.15, 1.5, 2.0, 2.5, 3.0],
                    value=1.5,
                    label="Line Spacing"
                )
                font_size = gr.Dropdown(
                    choices=["8", "10", "12", "14", "16", "18", "20", "22", "24"],
                    value="18",
                    label="Font Size"
                )
                alignment = gr.Dropdown(
                    choices=["Left", "Center", "Right", "Justified"],
                    value="Justified",
                    label="Text Alignment"
                )
                image_size = gr.Dropdown(
                    choices=["Small", "Medium", "Large"],
                    value="Small",
                    label="Image Size"
                )
                file_format = gr.Radio(["pdf", "docx"], label="File Format", value="pdf")
                get_document_btn = gr.Button(value="Get Document", elem_classes="download-btn")

        get_document_btn.click(
            generate_document, [input_media, output_text, file_format, font_size, line_spacing, alignment, image_size], gr.File(label="Download Document")
        )

demo.launch(debug=True)

Demo Output with ReportLab

Key Features

1. Visual Highlights Generator:

   • Extracts key objects, regions, and contextual clues from images and turns them into meaningful visual summaries.

2. Advanced Handwriting OCR:

   • Excels at recognizing and transcribing messy or cursive handwriting into digital text.

3. Vision-Language Fusion:

   • Seamlessly integrates visual input with language reasoning, ideal for image captioning, description, and Q&A.

4. Math and LaTeX Support:

   • Understands math problems in visual/text format and outputs in LaTeX syntax.

5. Conversational AI:

   • Supports multi-turn dialogue with memory of prior input — highly useful for interactive problem-solving and explanations.

6. Multi-modal Input Capability:

   • Accepts image, text, or a combination, and generates intelligent output tailored to the input.