CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning

📖Paper | 🏠Github |🤗CapRL-3B Model |🤗CapRL-InternVL3.5-8B Model | 🤗CapRL-2M Dataset

🤗CapRL Collection | 🤗Daily Paper ｜🤗CapRL-3B-GGUF ｜🤗CapRL-3B-i1-GGUF

CapRL-Eval-3B is the model used for answering questions based on captions, and it is a finetuned version of Qwen2.5-VL-3B. When dealing with tasks such as ChartQA (not multiple-choice questions), it provides more stable output formatting.

Introduction

We are excited to introduce CapRL-3B, a lightweight 3B image captioner that achieves perception capabilities comparable to Qwen2.5-VL-72B.

This is the first study of applying Reinforcement Learning with Verifiable Rewards for the open-ended and subjective image captioning task. Unlike traditional Supervised Fine-Tuning, which can lead to models memorizing a limited set of annotated captions, our method allows the model to explore and generate a broader range of creative and general descriptions. CapRL is a new training paradigm featuring a decoupled two-stage pipeline. The initial stage uses LVLMs to generate rich and accurate captions. Subsequently, the second stage evaluates caption quality by using a vision-only LLM to perform the QA task. We also created a specific QA curation pipeline to ensure the quality of the questions and answers used for the second stage.

By employing CapRL training framework, initializing with the Qwen2.5-VL-3B model, and using a carefully filtered 75K QA dataset as the training set, we obtained a highly capable captioner, CapRL-3B.

Key Features

• Remarkable visual understanding for Chart, Infographics and Document: CapRL-3B achieves perception accuracy and visual information coverage comparable to Qwen2.5-VL-72B.
• Well-organized output: The outputs of CapRL-3B are relatively well-structured, making them clear and easy to understand.
• Detailed description for natural images: The outputs of CapRL-3B can perfectly cover all valid visual information while containing fewer hallucinations.

Usage

If you want to use CapRL-3B for captioning, you can directly follow the exact same inference approach as in Qwen2.5-VL-series.

We recommend using vLLM to speed up inference.

Start an OpenAI API Service

Run the command below to start an OpenAI-compatible API service:

vllm serve "/PATH/CapRL-3B" \
    --trust-remote-code \
    --tensor-parallel-size=1 \
    --pipeline-parallel-size=1 \
    --gpu_memory_utilization=0.95 \
    --served-model-name=caprl \
    --port 8000 \
    --host 0.0.0.0

Then you can use the chat API as below: (see OpenAI API protocol document for more details):

import base64
from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"
client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)
image_path = "/path/to/local/image.png"
with open(image_path, "rb") as f:
    encoded_image = base64.b64encode(f.read())
encoded_image_text = encoded_image.decode("utf-8")
base64_qwen = f"data:image;base64,{encoded_image_text}"
chat_response = client.chat.completions.create(
    model="caprl",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": base64_qwen
                    },
                },
                {"type": "text", "text": "What is the text in the illustrate?"},
            ],
        },
    ],
    temperature=1.0,
    max_tokens=max_tokens,
    top_p=1.0,
    extra_body={
        "repetition_penalty": 1.0,
        },
)
print("Chat response:", chat_response)

Cases