---
license: apache-2.0
language:
- en
- zh
base_model:
- Qwen/Qwen2-VL-7B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- caption
- text-generation-inference
- flux
---
![9.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/esgwb8sdL5LbyDuQFLWnT.png)
# **JSONify-Flux-Large**

The **JSONify-Flux-Large** model is a fine-tuned version of **Qwen2VL**, specifically trained on **Flux-generated images** and their **corresponding captions**. This model has been trained using a **30M trainable parameter** dataset and is designed to output responses in structured **JSON format** while maintaining state-of-the-art performance in **Optical Character Recognition (OCR)**, **image-to-text conversion**, and **math problem-solving with LaTeX formatting**.

### Key Enhancements:

* **Optimized for Flux-Generated Image Captioning**: JSONify-Flux-Large has been trained to understand and describe images created using Flux-based generation techniques.

* **State-of-the-Art Image Understanding**: Built on Qwen2VL's architecture, JSONify-Flux-Large excels in visual reasoning tasks like DocVQA, RealWorldQA, MTVQA, and more.

* **Formatted JSON Output**: Responses are structured in a JSON format, making it ideal for automation, database storage, and further processing.

* **Multilingual Support**: Recognizes and extracts text from images in multiple languages, including English, Chinese, Japanese, Arabic, and various European languages.

* **Supports Multi-Turn Interactions**: Maintains context in conversations and can provide extended reasoning over multiple inputs.

### How to Use

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

# Load the model on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/JSONify-Flux-Large", torch_dtype="auto", device_map="auto"
)

# Enable flash_attention_2 for better acceleration and memory efficiency
# model = Qwen2VLForConditionalGeneration.from_pretrained(
#     "prithivMLmods/JSONify-Flux-Large",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# Default processor
processor = AutoProcessor.from_pretrained("prithivMLmods/JSONify-Flux-Large")

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 in JSON format."},
        ],
    }
]

# Prepare inputs 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")

# Inference: Generate JSON-formatted 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)  # JSON-formatted response
```

### JSON Buffer Handling
```python
buffer = ""
for new_text in streamer:
    buffer += new_text
    buffer = buffer.replace("<|im_end|>", "")
    yield buffer
```

### **Key Features**

1. **Flux-Based Vision-Language Model**:
   - Specifically trained on **Flux-generated images and captions** for precise image-to-text conversion.

2. **Optical Character Recognition (OCR)**:
   - Extracts and processes text from images with high accuracy.

3. **Math and LaTeX Support**:
   - Solves math problems and outputs equations in **LaTeX format**.

4. **Structured JSON Output**:
   - Ensures outputs are formatted in JSON, making it suitable for API responses and automation tasks.

5. **Multi-Image and Video Understanding**:
   - Supports analyzing multiple images and video content up to **20 minutes long**.

6. **Secure Weight Format**:
   - Uses **Safetensors** for enhanced security and faster model loading.