QvQ Step Tiny - [2B]

QvQ-Step-Tiny is a step-by-step context explainer Vision-Language model based on the Qwen2-VL architecture, fine-tuned using the VCR datasets for systematic step-by-step explanations. It is built on the Qwen2VLForConditionalGeneration framework with 2.21 billion parameters and uses BF16 (Brain Floating Point 16) precision.

Quickstart with Transformers

Here we show a code snippet to show you how to use the chat model with transformers and qwen_vl_utils:

from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

# default: Load the model on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/QvQ-Step-Tiny", torch_dtype="auto", device_map="auto"
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

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("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Key Enhancements of QvQ-Step-Tiny

1. State-of-the-Art Visual Understanding

   • QvQ-Step-Tiny inherits the state-of-the-art capabilities of Qwen2-VL for understanding images of various resolutions and aspect ratios.
   • It excels on visual reasoning benchmarks such as MathVista, DocVQA, RealWorldQA, and MTVQA, making it a powerful tool for detailed visual content analysis and question answering.

2. Extended Video Understanding

   • With the ability to process and comprehend videos of over 20 minutes, QvQ-Step-Tiny supports high-quality video-based question answering, conversational dialogs, and video content generation.
   • It ensures a systematic, step-by-step explanation of video content, which is ideal for educational, entertainment, and professional applications.

3. Integration with Devices and Systems

   • Thanks to its advanced reasoning and decision-making capabilities, QvQ-Step-Tiny can act as an intelligent agent for operating devices such as mobile phones, robots, and other automated systems.
   • It can process visual environments alongside textual instructions to enable seamless automation and intelligent control of devices.

4. Multilingual Support for Text in Images

   • QvQ-Step-Tiny supports multilingual text recognition within images, handling English, Chinese, and a wide range of languages, including most European languages, Japanese, Korean, Arabic, and Vietnamese.
   • This makes it an effective model for global applications, from document analysis to multi-language accessibility solutions.

Intended Use

1. Step-by-Step Context Explanation: Designed to provide detailed and systematic explanations for images and videos, making it ideal for educational, analytical, and instructional tasks.
2. Visual Content Understanding: Effective for analyzing visual content across diverse resolutions, aspect ratios, and formats, including documents (DocVQA) and mathematical visuals (MathVista).
3. Video-based Reasoning: Supports comprehension of long-form videos (20+ minutes) for tasks like video question answering, dialog generation, and instructional content creation.
4. Device Integration: Can act as an intelligent agent to automate device operations (e.g., mobile phones, robots) by understanding visual environments and processing text-based instructions.
5. Multilingual Visual Text Support: Recognizes and processes multilingual text within images, making it suitable for global applications like document processing and accessibility tools.
6. Advanced Question Answering: Excels in question-answering tasks involving images, videos, and multimodal data, serving as a robust tool for interactive systems.
7. Accessibility Enhancements: Assists visually impaired users by explaining visual and textual content in a clear, step-by-step manner.

Limitations

1. Model Size Constraints: At 2.21 billion parameters, it may not perform as well as larger models for highly complex or nuanced tasks.
2. Accuracy with Low-Quality Inputs: Performance may degrade when dealing with low-resolution images, poor lighting conditions, or noisy video/audio inputs.
3. Specialized Training Gaps: While strong on general benchmarks, it might struggle with niche or highly specialized domains that require additional fine-tuning.
4. Multilingual Text Variability: While multilingual text recognition is supported, performance may vary across less common or highly complex languages.
5. Context Length Tradeoffs: Processing very long videos (e.g., over 20 minutes) or highly dense visual data might challenge its coherence or explanation accuracy.
6. Device Integration Complexity: Deploying the model for operating devices or robots may require significant engineering efforts and robust integration pipelines.
7. Resource-Intensive for Long Contexts: Despite BF16 precision, tasks with extended context lengths or high-resolution inputs could demand substantial computational resources.
8. Ambiguity in Prompts: Ambiguously phrased or poorly structured input prompts may lead to incomplete or inaccurate explanations.
9. Static Model: The model cannot learn dynamically from user interactions or adapt its behavior without retraining.

Applications

• Education: Step-by-step explanations for visual and textual content in learning materials, including images and videos.
• Automation: Integrating with robotics or smart devices for performing tasks based on visual and textual data.
• Content Creation: Assisting in creating or analyzing video and image-based content, such as tutorials or product demos.
• Accessibility: Enhancing accessibility tools for visually impaired or multilingual users by providing clear explanations of image or video content.
• Global Q&A Systems: Supporting cross-lingual question answering in images and videos for diverse user bases.