metadata

license: apache-2.0
language:
  - en
  - zh
base_model:
  - Qwen/Qwen2-VL-2B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
tags:
  - text-generation-inference
  - label

Caption-Pro

Caption-Pro is an advanced image caption and annotation generator optimized for generating detailed, structured JSON outputs. Built upon a powerful vision-language architecture with enhanced OCR and multilingual support, Caption-Pro extracts high-quality captions and annotations from images for seamless integration into your applications.

Key Enhancements:

Advanced Image Understanding: Fine-tuned on millions of annotated images, Caption-Pro delivers precise comprehension and interpretation of visual content.
Optimized for JSON Output: Produces structured JSON data containing captions and detailed annotations—perfect for integration with databases, APIs, and automation pipelines.
Enhanced OCR Capabilities: Accurately extracts textual content from images in multiple languages, including English, Chinese, Japanese, Korean, Arabic, and more.
Multimodal Processing: Seamlessly handles both image and text inputs, generating comprehensive annotations based on the provided image.
Multilingual Support: Recognizes and processes text within images across various languages.
Secure and Optimized Model Weights: Employs safetensors for efficient and secure model loading.

How to Use

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

# Load the Caption-Pro model with optimized parameters
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/Caption-Pro", torch_dtype="auto", device_map="auto"
)

# Recommended acceleration for performance optimization:
# model = Qwen2VLForConditionalGeneration.from_pretrained(
#     "prithivMLmods/Caption-Pro",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# Load the default processor for Caption-Pro
processor = AutoProcessor.from_pretrained("prithivMLmods/Caption-Pro")

# Define the input messages with both an image and a text prompt
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://flux-generated.com/sample_image.jpeg",
            },
            {"type": "text", "text": "Provide detailed captions and annotations for this image in JSON format."},
        ],
    }
]

# Prepare the input for inference
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")

# Generate the output
generated_ids = model.generate(**inputs, max_new_tokens=256)
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 Features

Annotation-Ready Training Data
- Trained using a diverse dataset of annotated images to ensure high-quality structured output.
Optical Character Recognition (OCR)
- Robustly extracts and processes text from images in various languages and scripts.
Structured JSON Output
- Generates detailed captions and annotations in standardized JSON format for easy downstream integration.
Image & Text Processing
- Capable of handling both visual and textual inputs, delivering comprehensive and context-aware annotations.
Conversational Annotation Generation
- Supports multi-turn interactions, enabling detailed and iterative refinement of annotations.
Secure and Efficient Model Weights
- Uses safetensors for enhanced security and optimized model performance.

Caption-Pro streamlines the process of generating image captions and annotations, making it an ideal solution for applications that require detailed visual content analysis and structured data integration.