Image-Text-to-Text

Different Types of Vision Language Models

Vision language models come in three types:

• Base: Pre-trained models that can be fine-tuned. A good example of base models is the PaliGemma models family by Google.
• Instruction: Base models fine-tuned on instruction datasets. A good example of instruction fine-tuned models is idefics2-8b.
• Chatty/Conversational: Base models fine-tuned on conversation datasets. A good example of chatty models is deepseek-vl-7b-chat.

Use Cases

Multimodal Dialogue

Vision language models can be used as multimodal assistants, keeping context about the conversation and keeping the image to have multiple-turn dialogues.

Zero-shot Object Detection, Image Segmentation and Localization

Some vision language models can detect or segment a set of objects or describe the positions or relative positions of the objects. For example, one could prompt such a model to ask if one object is behind another. Such a model can also output bounding box coordination or segmentation masks directly in the text output, unlike the traditional models explicitly trained on only object detection or image segmentation.

Visual Question Answering

Vision language models trained on image-text pairs can be used for visual question answering and generating captions for images.

Document Question Answering and Retrieval

Documents often consist of different layouts, charts, tables, images, and more. Vision language models trained on formatted documents can extract information from them. This is an OCR-free approach; the inputs skip OCR, and documents are directly fed to vision language models.

Image Recognition with Instructions

Vision language models can recognize images through descriptions. When given detailed descriptions of specific entities, it can classify the entities in an image.

Inference

You can use the Transformers library to interact with vision-language models. Specifically, pipeline makes it easy to infer models.

Initialize the pipeline first.

from transformers import pipeline

pipe = pipeline("image-text-to-text", model="llava-hf/llava-interleave-qwen-0.5b-hf")

The model's built-in chat template will be used to format the conversational input. We can pass the image as an URL in the content part of the user message:

messages = [
     {
         "role": "user",
         "content": [
             {
                 "type": "image",
                 "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
             },
             {"type": "text", "text": "Describe this image."},
         ],
     }
 ]

We can now directly pass in the messages to the pipeline to infer. The return_full_text flag is used to return the full prompt in the response, including the user input. Here we pass False to only return the generated text.

outputs = pipe(text=messages, max_new_tokens=60, return_full_text=False)

outputs[0]["generated_text"]
# The image captures a moment of tranquility in nature. At the center of the frame, a pink flower with a yellow center is in full bloom. The flower is surrounded by a cluster of red flowers, their vibrant color contrasting with the pink of the flower. \n\nA black and yellow bee is per

You can also use the Inference API to test image-text-to-text models. You need to use a Hugging Face token for authentication.

curl https://api-inference.huggingface.co/models/meta-llama/Llama-3.2-11B-Vision-Instruct \
    -X POST \
    -d '{"messages": [{"role": "user","content": [{"type": "image"}, {"type": "text", "text": "Can you describe the image?"}]}]}' \
    -H "Content-Type: application/json" \
    -H "Authorization: Bearer hf_***"

Useful Resources

• Vision Language Models Explained
• Open-source Multimodality and How to Achieve it using Hugging Face
• Introducing Idefics2: A Powerful 8B Vision-Language Model for the community
• Image-text-to-text task guide
• Preference Optimization for Vision Language Models with TRL