metadata

license: apache-2.0

R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcement Learning

📖 Introduction

R1-Omni is the industry’s first application of Reinforcement Learning with Verifiable Reward (RLVR) to an Omni-multimodal large language model. We focus on emotion recognition, a task where both visual and audio modalities play crucial roles, to validate the potential of combining RLVR with Omni model. Our findings reveal several key insights:

Enhanced Reasoning Capability: R1-Omni demonstrate superior reasoning abilities, enabling a clearer understanding of how visual and audio information contribute to emotion recognition.
Improved Understanding Capability: Compared to SFT, RLVR significantly boosts performance on emotion recognition tasks.
Stronger Generalization Capability: RLVR models exhibit markedly better generalization capabilities, particularly excelling in out-of-distribution scenarios.

🏆 Performance

Below are the performance on emotion recognition datasets. We use symbols to indicate whether the data is in-distribution (⬤) or out-of-distribution (△).

Method	DFEW (WAR) ⬤	DFEW (UAR) ⬤	MAFW (WAR) ⬤	MAFW (UAR) ⬤	RAVDESS (WAR) △	RAVDESS (UAR) △
HumanOmni-0.5B	22.64	19.44	20.18	13.52	7.33	9.38
EMER-SFT	38.66	35.31	38.39	28.02	29.00	27.19
MAFW-DFEW-SFT	60.23	44.39	50.44	30.39	29.33	30.75
R1-Omni	65.83	56.27	57.68	40.04	43.00	44.69

Legend

⬤: Indicates in-distribution data (DFEW and MAFW).
△: Indicates out-of-distribution data (RAVDESS).

🛠️ Environment Setup

Our code is built on the R1-V framework. To set up the environment, please follow the installation instructions in the R1-V repository

🔍 Inference

Our inference code is based on the implementation from HumanOmni. To ensure the model runs inference smoothly, follow these steps:

Download the Required Models：
- siglip-224.
- whisper-large-v3.
Update the Configuration File：
- In the directory where you downloaded the R1-Omni model, locate the config.json file.
- Update the paths on line 23 and line 31 to point to the local folders where you saved the models.

Example: Updating config.json

If you saved the models to the following local paths:：

/path/to/local/models/siglip-base-patch16-224
/path/to/local/models/whisper-large-v3

Update the relevant lines in config.json as follows：

 "mm_audio_tower": "/path/to/local/models/whisper-large-v3",
 "mm_vision_tower": "/path/to/local/models/siglip-base-patch16-224"

We provide inference.py for singe video inference.

python inference.py --modal video_audio \
  --model_path ./R1-Omni-0.5B \
  --video_path video.mp4 \
  --instruct "As an emotional recognition expert; throughout the video, which emotion conveyed by the characters is the most obvious to you?  Output the thinking process in <think> </think> and final emotion in <answer> </answer> tags."

🧠 Training

Cold Start

we initialize the HumanOmni-0.5B by fine-tuning it on a combined dataset consisting of 232 samples from the Explainable Multimodal Emotion Reasoning dataset and 348 samples from HumanOmni dataset. An example json file of the training data:

[
    {
        "video": "MER24/sample_00000967.mp4",
        "conversations": [
            {
                "from": "human",
                "value": "<video>\n<audio>\nAs an emotional recognition expert; throughout the video, which emotion conveyed by the characters is the most obvious to you? Output the thinking process in <think> </think> and final emotion in <answer> </answer> tags."
            },
            {
                "from": "gpt",
                "value": "<think>The video depicts a bright and tranquil indoor setting, where a man in a white Polo shirt stands by the window, engaged in a phone call. His furrowed brow and open mouth suggest he is experiencing tension and anxiety. According to the audio content of the video, his speech is fast-paced, and his tone is filled with confusion and stress. A comprehensive analysis reveals that the man is facing a moderate level of anxiety, closely linked to the challenging phone conversation he is having. Consequently, the entire emotional analysis report emphasizes his anxiety and nervousness in handling challenging situations.</think>\n<answer>anxious</answer>"
            }
        ]
    },
  ...
]

All of the cold-start data will be released as soon as possible.

RLVR

In this stage, we utilize the training sets from the MAFW and DFEW(https://dfew-dataset.github.io/) datasets, comprising a total of 15,306 video sample. An example json file of the training data:

[
    {
        "video": "DFEW/videos/1.mp4",
        "conversations": [
            {
                "from": "human",
                "value": "<video>\n<audio>\nAs an emotional recognition expert; throughout the video, which emotion conveyed by the characters is the most obvious to you?"
            },
            {
                "from": "gpt",
                "value": "sad"
            }
        ],
    },
  ...
]

🤝 Related Work

📚 Citation

If you find our work helpful, feel free to give us a cite.

{zhao2025r1omniexplainableomnimultimodalemotion,
      title={R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcement Learning}, 
      author={Jiaxing Zhao and Xihan Wei and Liefeng Bo},
      journal={arXiv preprint arXiv:2503.05379},
      year={2025}
}