<h1 align="center">(ICLR 2026)CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning</h1> <a href="https://github.com/Cooperx521">Long Xing*</a> · <a href="https://lightdxy.github.io/">Xiaoyi Dong*</a> · <a href="https://yuhangzang.github.io/">Yuhang Zang</a> · <a href="https://scholar.google.com/citations?user=sJkqsqkAAAAJ">Yuhang Cao</a> · <a href="https://scholar.google.com/citations?user=P4yNnSkAAAAJ&hl=zh-TW">Jianze Liang</a> · <a href="https://github.com/shikiw">Qidong Huang</a> · <a href="https://myownskyw7.github.io/">Jiaqi Wang</a> · <a href="https://scholar.google.com/citations?user=5bInRDEAAAAJ&hl=zh-CN">Feng Wu</a> · <a href="http://dahua.site/">Dahua Lin</a>

📖<a href="https://arxiv.org/abs/2509.22647">Paper</a> | 🏠<a href="https://github.com/InternLM/CapRL">Github</a> | 🤗<a href="https://huggingface.co/collections/long-xing1/caprl-68d64ac32ded31596c36e189">CapRL Collection</a> | 🤗<a href="https://huggingface.co/papers/2509.22647">Daily Paper</a>

CapRL Series Model & Dataset

We are excited to release the CapRL 2.0 series: CapRL-Qwen3VL-2B and CapRL-Qwen3VL-4B. These models feature fewer parameters while delivering even more powerful captioning performance. Notably, CapRL-Qwen3VL-2B outperforms both CapRL-Qwen2.5VL-3B and Qwen2.5VL-72B in captioning tasks, while CapRL-Qwen3VL-4B further demonstrates a significant performance leap over the 2B version. This improvement in efficiency is driven by our upgraded training recipe, which includes a more rigorous QA data filter and a significantly more diverse image dataset. We welcome everyone to try them out!

Now you can try out CapRL with your own images🎨! ➡️ 🌈CapRL-Qwen2.5VL-3B Space and 🌈CapRL-Qwen3VL-4B Space.

📢 News

We are working on even stronger base models and upgrading our training recipe — stay tuned!

🔥 [2/9/2026] We release the CapRL training code.
🔥 [1/27/2026] CapRL is accepted by ICLR2026! We are working on cleaning training code, and will release everything as soon as possible!
🔥 [12/24/2025] We are excited to release the CapRL 2.0 series: CapRL-Qwen3VL-2B and CapRL-Qwen3VL-4B!
🔥 [12/24/2025] The total downloads of the CapRL-related models and dataset reached 17,000!
🔥 [10/15/2025] The total downloads of the CapRL-related models and dataset reached 6,000 within just 20 days!
🚀 [10/15/2025] We are excited to announce the release of CapRL-InternVL3.5-8B, whose image captioning capability outperforms Qwen2.5-VL-72B!
🚀 [10/15/2025] Thanks mradermacher for the valuable contribution! CapRL-3B-GGUF is the static quants version, and CapRL-3B-i1-GGUF is weighted/imatrix quants version.
🚀 [10/15/2025] We release QA curation code.
🚀 [09/25/2025] We release CapRL repository, CapRL-3B model, evaluation code and dataset.

Introduction

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

💡 Highlights

🔥 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.

Model Card

Based on the same recipe as CapRL-3B, we used InternVL3.5-8B as the policy model and obtained CapRL-InternVL3.5-8B through CapRL.
CapRL-3B-GGUF is static quants version, and CapRL-3B-i1-GGUF is weighted/imatrix quants version. Thanks for their contribution!

👨‍💻 Todo

[ ] Release 75k QA dataset.

🛠️ Setup

Installation

git clone https://github.com/InternLM/CapRL.git
cd CapRL/CapRL_Training
conda create -n CapRL python=3.10
conda activate CapRL
bash setup.sh

The setup.sh will sequentially:

Install key dependencies with pinned versions (torch, transformers, vllm, deepspeed, flash-attn, ray, etc.)
Install the OpenRLHF-based training framework and remaining dependencies via pip install -e .

⭐️ Quick Start

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

The prompt we use for training and evaluation is Please describe this image in detail.

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": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": base64_qwen
                    },
                },
                {"type": "text", "text": "Please describe this image in detail."},
            ],
        },
    ],
    temperature=1.0,
    max_tokens=max_tokens,
    top_p=1.0,
    extra_body={
        "repetition_penalty": 1.0,
        },
)
print("Chat response:", chat_response)

QA Curation

This part of the code is in the QA_data_curation folder, which contains all four steps for generating QA data:

QA generation. Use Qwen2.5-VL-72B to generate 5 QAs for each image. The generation process launches a vLLM service and uses multi-threading to speed up.
QA extraction. Extract QAs through format matching.
Qwen2.5-VL-3B answer question. Use Qwen2.5-VL-3B to answer questions with and without images. The parameter ROTATE_NUM controls how many times each question is answered. If a question is answered only once, the randomness may be too high and can easily lead to misjudgment.
Filter question. We keep QA pairs with visual acc higher than 0.75 and text acc lower than 0.25 to avoid data leakage and ensure the model can correctly answer questions when images are provided.

CapRL

Install / Use

README