Thinking with Images

Use image zoom and search tools for precise comprehension of fine-grained visual details

Qwen3-VL can utilize specialized tools like image_zoom_in and search_tool to facilitate precise comprehension of fine-grained visual details within images. This enables deeper visual reasoning and analysis.

Capability Overview

The thinking with images feature enables you to:

• Zoom into specific image regions for detail analysis
• Use search tools to find specific visual elements
• Perform fine-grained visual reasoning
• Analyze complex visual details
• Combine multiple visual analysis tools
• Enable step-by-step visual problem solving

Example Usage

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "Qwen/Qwen3-VL-235B-A22B-Thinking", dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-235B-A22B-Thinking")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "path/to/complex_image.jpg",
            },
            {"type": "text", "text": "Analyze the fine details in this image and explain what you observe."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
)
inputs = inputs.to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=1024)
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)

Try it Yourself

Explore the full thinking with images cookbook with interactive examples:

View on GitHub