InteractiveOmni

InteractiveOmni-4B 🤗  | InteractiveOmni-8B 🤗  | 📑 Paper   

Introduction

InteractiveOmni is a unified omni-modal model that can simultaneously receive inputs such as images, audio, text, and video and directly generate coherent text and speech streams, achieving truly integrated interaction.

This is the schematic diagram for multi-turn audio-visual interaction.

Key Features

• Strong Performance Across Modalities: Exhibiting omni-modal understanding and speech generation capabilities. InteractiveOmni outperforms the similarly sized vision-language models, audio-language models and omni-modal models.
• State-of-the-Art Performance: Achieve SOTA results on various open-source benchmarks for image, audio, and video understanding, as well as speech conversation.
• Excellent Interactive Performance: Achieve more intelligent audio-visual experience with multi-turn and long-term memory capabilities.
• Multi-turn Interactive Benchmarks: Propose multi-modal multi-turn benchmark to evaluate multi-turn memory and speech interaction of leading MLLMs.
• On-device Model: the 4B model achieves 97% of the performance with just 50% of the model size compared with 8B model.

Model Architecture

Quickstart

Get the Code

git clone https://github.com/OpenSenseNova/InteractiveOmni.git
cd InteractiveOmni
pip install -r requirements.txt

We provide an example code to run InteractiveOmni using 🤗 Transformers.

Please use transformers>=4.51.0 and FlashAttention2 to ensure the model works normally.

Model Loading

import torch
from transformers import AutoTokenizer, AutoModel
path = "sensenova/InteractiveOmni-8B"
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True).eval().cuda()

Inference with Transformers

import torch
from transformers import AutoModel, AutoTokenizer
import torchaudio

path = "sensenova/InteractiveOmni-8B"
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=True)

# set the max number of tiles in `max_num`
max_num = 12
frame = 8
generation_config = dict(max_new_tokens=1024, do_sample=True)

# pure-text conversation (纯文本对话)
messages = [
    {
        'role': "user",
        'content': 'Hello, who are you?',
    }
]
response = model.chat(tokenizer, generation_config, messages)

# audio conversation (音频对话)
messages = [
    {
        'role': "user",
        'content': [
            {
                "type": "audio",
                "audio": "assets/hello_en.wav"
            }
        ]
    }
]
response = model.chat(tokenizer, generation_config, messages)

## Generate both audio and text output
messages = [
    {
        'role': "user",
        'content': [
            {
                "type": "audio",
                "audio": "assets/hello_zh.wav"
            }
        ]
    }
]
response, wav_response = model.chat(tokenizer, generation_config, messages, generate_audio=True)
torchaudio.save("result.wav", wav_response.cpu(), 24000, format="wav")

# image-text conversation (图文对话)
messages = [
    {
        'role': "user",
        'content': [
            {
                "type": "image",
                "image": 'assets/cat_cup.jpeg'
            },
            {
                "type": "text",
                "text": "Please describe the image shortly."
            }
        ]
    }
]
response = model.chat(tokenizer, generation_config, messages, max_num)

# image-audio conversation (图音对话)
messages = [
    {
        'role': "user",
        'content': [
            {
                "type": "image",
                "image": 'assets/cat_cup.jpeg'
            },
            {
                "type": "audio",
                "audio": "assets/describe_img_en.wav"
            }
        ]
    }
]
response = model.chat(tokenizer, generation_config, messages, max_num)

## image-audio conversation, generate both audio and text output
messages = [
    {
        'role': "user",
        'content': [
            {
                "type": "image",
                "image": 'assets/cat_cup.jpeg'
            },
            {
                "type": "audio",
                "audio": "assets/describe_img_en.wav"
            }
        ]
    }
]
response, wav_response = model.chat(tokenizer, generation_config, messages, generate_audio=True)
torchaudio.save("result.wav", wav_response.cpu(), 24000, format="wav")

# video conversation (视频对话)
messages = [
    {
        'role': "user",
        'content': [
            {
                "type": "video",
                "video": 'video_path'
            },
            {
                "type": "text",
                "text": "Describe this video in detail."
            }
        ]
    }
]
response = model.chat(tokenizer, generation_config, messages, max_num, frame)

Use audio output

• If users need audio output, the system prompt must be set as follows, otherwise the audio output may not work as expected.

You are a highly advanced multimodal conversational AI designed for human-like interaction. You can perceive auditory, visual, speech, and textual inputs, and generate text and speech.

messages = [
    {
        "role": "system",
        "content": "You are a highly advanced multimodal conversational AI designed for human-like interaction. You can perceive auditory, visual, speech, and textual inputs, and generate text and speech."
    },
    {
        'role': "user",
        'content': [
            {
                "type": "audio",
                "audio": "assets/hello_zh.wav",
            }
        ]
    }
]
response, wav_response = model.chat(tokenizer, generation_config, messages, generate_audio=True)
torchaudio.save("result_none_speaker.wav", wav_response.cpu(), 24000, format="wav")

• Use default speaker to generate output audio.

messages = [
    {
        "role": "system",
        "content": "You are a highly advanced multimodal conversational AI designed for human-like interaction. You can perceive auditory, visual, speech, and textual inputs, and generate text and speech."
    },
    {
        'role': "user",
        'content': [
            {
                "type": "audio",
                "audio": "assets/hello_zh.wav",
            }
        ]
    }
]
response, wav_response = model.chat(tokenizer, generation_config, messages, generate_audio=True, speaker_embedding=model.default_speaker_embedding)
torchaudio.save("result_default_speaker.wav", wav_response.cpu(), 24000, format="wav")

• Use custom speaker to generate output audio, similar to sound cloning.

messages = [
    {
        "role": "system",
        "content": "You are a highly advanced multimodal conversational AI designed for human-like interaction. You can perceive auditory, visual, speech, and textual inputs, and generate text and speech."
    },
    {
        'role': "user",
        'content': [
            {
                "type": "audio",
                "audio": "assets/hello_zh.wav",
            }
        ]
    }
]
speaker_embedding = model.extract_speaker_embedding("assets/hello_zh.wav")
response, wav_response = model.chat(tokenizer, generation_config, messages, generate_audio=True, speaker_embedding=speaker_embedding)
torchaudio.save("result_custom_speaker.wav", wav_response.cpu(), 24000, format="wav")

Evaluation

InteractiveOmni achieves state-of-the-art performance across a wide range of multi-modal understanding and speech generation benchmarks.

Image Understanding Video Understanding Audio Understanding Omni-modal UnderstandingSpeech-to-text Speech Generation

Citation

If you find our paper and code useful in your research, please cite our technical report.

@misc{tong2025interactiveomniunifiedomnimodalmodel,
      title={InteractiveOmni: A Unified Omni-modal Model for Audio-Visual Multi-turn Dialogue}, 
      author={Wenwen Tong and Hewei Guo and Dongchuan Ran and Jiangnan Chen and Jiefan Lu and Kaibin Wang and Keqiang Li and Xiaoxu Zhu and Jiakui Li and Kehan Li and Xueheng Li and Lumin Li and Chenxu Guo and Jiasheng Zhou and Jiandong Chen and Xianye Wu and Jiahao Wang and Silei Wu and Lei Chen and Hanming Deng and Yuxuan Song and Dinghao Zhou and Guiping Zhong and Ken Zheng and Shiyin Kang and Lewei Lu},
      year={2025},
      eprint={2510.13747},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.13747}, 
}