HumanOmni: A Large Vision-Speech Language Model for Human-Centric Video Understanding

News

1. Building upon HumanOmni, we are the first to combine RLVR (Reinforcement Learning for Vision and Reasoning) with an Omni model, introducing R1-Omni, a reasoning-based large multimodal model.

📖 Introduction

HumanOmni is the industry’s first human-centric Omni-multimodal large language model for comprehensive understanding in human-centric scenes.

1. 2.4M human-centric video clips with over 14M double-check instructions: We have constructed a dataset containing over 2.4M human-centric video clips, providing rich and detailed information about individuals. We provide over 14M instruction data for visual pretraining.
2. 50K video clips with more than 100K manually annotated instrcutions: We have manually annotated 50K video clips with more than 100K instructions related to emotion recognition, facial description, and speaker-specific speech recognition for visual fine-tuning and cross-modal interaction integration.
3. Three human-specific branch: We use three branches to handle face-related, body-related, and interaction-related scenes separately in HumanOmni. HumanOmni dynamically adjusts its fusion weights based on input instructions, ensuring accurate responses across various scenes.
4. Audio-visual synergy: HumanOmni can simultaneously understand vision and speech, allowing for a more comprehensive understanding of complex scenes.

📦 Model Download

Model	Stage	#Params	HuggingFace	ModelScope
HumanOmni-Video	Visual Capability Construction	7B
HumanOmni-Audio	Auditory Capability Development	7B
HumanOmni-Omni	Cross-Modal Interaction Integration	7B

Our training pipeline consists of three progressive stages to establish multimodal understanding capabilities:

📹 Visual Capability Construction

• Model: HumanOmni-Video
• Objective: Learn spatio-temporal feature representations to analyze human actions and scene dynamics in videos.

🎧 Auditory Capability Development

• Model: HumanOmni-Audio
• Objective: Develop robust speech comprehension and audio interpretation through large-scale acoustic modeling.

🌐 Cross-Modal Interaction Integration

• Model: HumanOmni-Omni (also referred to as HumanOmni)
• Objective: Enable synergistic vision-audio reasoning by fine-tuning parameters from both HumanOmni-Video and HumanOmni-Audio.

🏆 Performance

• Emotion Understanding:

Method	Modalities	DFEW (UAR)	DFEW (WAR)	MAFW (UAR)	MAFW (WAR)
Specialized models for emotion-related tasks
Wav2Vec2.0	A	36.15	43.05	21.59	29.69
HuBERT	A	35.98	43.24	25.00	32.60
DFER-CLIP	V	59.61	71.25	38.89	52.55
MAE-DFER	V	63.41	74.43	41.62	54.31
HiCMAE	AV	63.76	75.01	42.65	56.17
Emotion-LLaMA	AV	64.21	77.06	-	-
MMA-DFER	AV	66.85	77.43	44.25	58.45
Other models
Qwen2-VL-7B	V	43.08	52.83	31.67	45.89
Qwen2-VL-72B	V	39.24	45.12	42.61	46.07
VITA	AV	21.36	32.07	14.05	33.38
InternLM-XComposer-2.5-OL	AV	44.23	51.29	33.78	46.81
GPT4-O	AV	50.57	57.19	38.29	48.82
HumanOmni	AV	74.86	82.46	52.94	68.40

• Dynamic Facial Expression Caption:

Method	Correctness	Detail	Context	Temporal	CIDEr	Rouge-L	AutoDQ
Vision large language model
VideoLLaMA	3.60	3.67	3.84	3.50	0.189	0.196	0.303
VideoChat	3.47	3.52	3.92	3.38	0.251	0.192	0.344
VideoChat2	3.70	3.56	4.16	3.52	0.202	0.229	0.311
Chat-UniVI	3.64	3.63	4.21	3.61	0.189	0.231	0.396
LLaVA-Next-Video	4.19	4.07	4.39	4.04	0.250	0.249	0.395
ShareGPT4Video	4.24	4.13	4.35	4.09	0.192	0.205	0.394
LLaMA-VID	3.95	4.01	4.22	3.71	0.195	0.231	0.339
VideoLLaMA2	4.17	4.02	4.47	3.93	0.253	0.266	0.344
PLLaVA	4.21	4.15	4.37	4.08	0.268	0.250	0.393
ST-LLM	4.00	3.98	4.31	3.94	0.213	0.238	0.321
Tarsier	3.59	3.50	4.07	3.41	0.143	0.185	0.415
LLaVA-OneVision	3.68	3.47	4.10	3.42	0.115	0.165	0.379
FaceTrack-MM	4.42	4.30	4.60	4.26	0.418	0.473	0.483
Qwen2-VL-72B	4.28	4.14	4.55	4.08	0.241	0.314	0.449
Qwen2-VL-7B	4.23	4.16	4.52	4.02	0.204	0.233	0.422
Qwen2-VL-2B	4.01	3.98	4.37	3.88	0.202	0.221	0.406
Claude3.5-Sonnet	4.13	4.01	4.49	4.05	0.243	0.228	0.442
Omni-modality large language model
GPT4-O	4.22	3.97	4.48	3.90	0.264	0.213	0.432
VITA	3.98	3.74	4.11	3.59	0.191	0.224	0.366
InternLM-XComposer-2.5-OL	3.91	3.70	4.12	3.54	0.113	0.164	0.382
HumanOmni	4.58	4.41	4.70	4.41	0.412	0.468	0.523

• Action and Pose Understanding:

Method	Action Sequence	Unexpected Action	Action Antonym	Object Interaction	Action Count	Fine-grained Action	Avg
Vision large language model
Otter-V	23.0	29.5	27.5	28.0	26.0	27.0	26.8
mPLUG-Owl-V	22.0	29.0	34.0	27.0	31.5	29.0	28.8
Video-LLaMA	27.5	39.0	51.0	40.5	34.0	29.0	36.8
LLaMA-Adapter	23.0	33.0	51.0	32.5	29.0	30.0	33.1
Video-ChatGPT	23.5	26.5	62.0	28.0	30.5	22.5	32.2
VideoChat	33.5	40.5	56.0	40.5	35.0	33.5	39.8
VideoChat2	75.5	60.5	83.5	74.5	37.0	50.5	63.6
ST-LLM	66.0	58.5	84.0	73.5	36.5	44.0	60.4
PLLaVA	58.0	61.0	55.5	61.0	39.5	41.0	52.6
VideoLLaMB	54.5	52.0	86.5	58.5	40.5	44.5	56.1
Qwen2-VL-72B*	51.5	82.0	93.5	81.5	48.5	49.0	67.7
Qwen2-VL-7B*	73.5	80.0	79.0	78.5	46.0	49.0	67.7
Qwen2-VL-2B*	77.5	76.5	76.5	77.5	50.0	47.5	67.6
GPT-4V	55.5	63.5	72.0	59.0	39.0	47.5	56.1
Omni-modality large language model
VITA	58.0	81.5	73.5	61.5	45.5	42.0	60.3
InternLM-XComposer-2.5-OL	84.5	81.0	75.0	79.5	60.5	46.0	71.1
HumanOmni	70.0	78.0	92.5	80.5	65.5	49.0	72.6

📺 Demo

00772.mp4

User:What's the major emotion?
HumanOmni:angry.

User:What's the major facial expression?
HumanOmni:frown.

User:Desribe the woman.
HumanOmni:The woman has long brown hair and is wearing a pink top with a dark cardigan.

User:Please describe the major action in detail.
HumanOmni:The woman is frowning and shaking her head, indicating that she is angry. She is also gesturing with her hands as if she is explaining something or emphasizing a point.

User:What did the woman say?
HumanOmni:The woman said "Of course we object! I object, Steve!"

01-01-03-01-02-02-08.mp4

User:What's the major emotion?
HumanOmni:happy.

User:What's the major facial expression?
HumanOmni:smile.

User:Desribe the woman in detail.
HumanOmni:The woman in the video has blonde hair, is wearing a black top, and is smiling.

User:What did the woman say?
HumanOmni:The woman said "Dogs are sitting by the door."

chinese_video.mp4

User:描述一下视频中男子的外貌穿着.
HumanOmni:视频中的男性有短的黑发，配戴圆框眼镜。他穿了一件棕色的毛大衣，内里是白色的.

User:视频中男子说了什么内容？
HumanOmni:男子说他穿什么颜色的衣服，现在处于什么环境，今天的天气怎么样.

User:视频中男子有什么情绪吗？
HumanOmni:没有情绪。

🛠️ Environment Setup

To set up the recommended environment for HumanOmni, follow these instructions:

Recommended Environment

• Python: >=3.10
• CUDA: >=12.1
• PyTorch: >=2.2 (with CUDA support)
• Transformers: >=4.45
• Accelerate: >=0.30.1

Or you can quickly set up the environment as follows:

git clone https://github.com/HumanMLLM/HumanOmni
cd HumanOmni
conda create -n humanOmni python=3.10 -y
conda activate humanOmni
pip install --upgrade pip
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

🧠 Training on Custom Dataset

Data Preparation

An example json file of the training data:

[
    {
        "video": "human/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?\nfear ,angry ,surprise ,happy ,neutral ,sad ,disgust"
            },
            {
                "from": "gpt",
                "value": "sad"
            }
        ],
    },
    {
        "video": "human/DFEW/videos/1.mp4",
        "conversations": [
            {
                "from": "human",
                "value": "<video>\n<audio>\nAs an emotional recognition expert, in the video, when the characters display their emotions, which predominant feeling is most clearly expressed?\nfear ,disgust ,happy ,sad ,surprise"
            },
            {
                "from": "gpt",
                "value": "sad"
            }
        ],
    },
  ...
]

Multi-Modal SFT

• Download the required weights: (1) HumanOmni-7B-Video (2) HumanOmni-7B-Audio
• scripts/train/finetune_humanomni.sh Loading the weights and the prepared dataset.
• bash scripts/train/finetune_humanomni.sh

🔍 Inference

We provide inference.py for singe video inference.

• video + audio

python inference.py --modal video_audio \
  --model_path ./HumanOmni_7B \
  --video_path video.mp4 \
  --instruct "Describe this video."

• only video

python inference.py --modal video \
  --model_path ./HumanOmni_7B \
  --video_path video.mp4 \
  --instruct "Describe this video."

• only audio

python inference.py --modal audio \
  --model_path ./HumanOmni_7B \
  --video_path video.mp4 \
  --instruct "Describe this video."

🤝 Related Work

• LLaVA-Octopus: Unlocking Instruction-Driven Adaptive Projector Fusion for Video Understanding
• Omni-Emotion: Extending Video MLLM with Detailed Face and Audio Modeling for Multimodal Emotion Analysis
• Qwen2.5

📚 Citation

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

@article{zhao2025humanomni,
  title={HumanOmni: A Large Vision-Speech Language Model for Human-Centric Video Understanding},
  author={Zhao, Jiaxing and Yang, Qize and Peng, Yixing and Bai, Detao and Yao, Shimin and Sun, Boyuan and Chen, Xiang and Fu, Shenghao and Wei, Xihan and Bo, Liefeng and others},
  journal={arXiv preprint arXiv:2501.15111},
  year={2025}
}