File size: 6,193 Bytes
c48c32c 7b3777b c48c32c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 | ---
library_name: transformers
license: apache-2.0
tags:
- omni-modal
- multimodal
- vision
- audio
- video
- llm
model-index:
- name: OmniVinci
results:
- task:
type: image-to-text
name: Image Understanding
dataset:
name: MVBench
type: mvbench
metrics:
- name: MVBench Score
type: accuracy
value: 70.6
source:
name: OmniVinci Technical Report
url: https://arxiv.org/abs/2510.15870
- task:
type: video-to-text
name: Video Understanding
dataset:
name: Video-MME
type: video-mme
metrics:
- name: Video-MME (w/o sub)
type: accuracy
value: 68.2
source:
name: OmniVinci Technical Report
url: https://arxiv.org/abs/2510.15870
- task:
type: video-to-text
name: Cross-Modal Understanding
dataset:
name: DailyOmni
type: dailyomni
metrics:
- name: DailyOmni Score
type: accuracy
value: 66.5
source:
name: OmniVinci Technical Report
url: https://arxiv.org/abs/2510.15870
- task:
type: audio-to-text
name: Audio Understanding
dataset:
name: MMAR
type: mmar
metrics:
- name: MMAR Score
type: accuracy
value: 58.4
source:
name: OmniVinci Technical Report
url: https://arxiv.org/abs/2510.15870
- task:
type: audio-to-text
name: Audio-Only Reasoning
dataset:
name: MMAU
type: mmau
metrics:
- name: MMAU Score
type: accuracy
value: 71.6
source:
name: OmniVinci Technical Report
url: https://arxiv.org/abs/2510.15870
- task:
type: video-to-text
name: Multi-Modal Reasoning
dataset:
name: Worldsense
type: worldsense
metrics:
- name: Worldsense Score
type: accuracy
value: 48.2
source:
name: OmniVinci Technical Report
url: https://arxiv.org/abs/2510.15870
---
# <span style="background: linear-gradient(45deg, #667eea 0%, #764ba2 25%, #f093fb 50%, #f5576c 75%, #4facfe 100%); -webkit-background-clip: text; -webkit-text-fill-color: transparent; background-clip: text; font-weight: bold; font-size: 1.1em;">**OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM**</span> <br />
[](https://arxiv.org/abs/2510.15870)
[](https://github.com/NVlabs/OmniVinci)
[](https://huggingface.co/nvidia/omnivinci)
[](https://nvlabs.github.io/OmniVinci)
## Introduction
OmniVinci is an NVIDIA research project focused on exploring omni-modal LLMs that can not only see and read but also listen, speak, and reason.
We are among the best omni-modality understanding models. Check out our performance on some of the most popular omni-modality, audio, and vision benchmarks:
<p align="center">
<img src="./asset/performance.png" width="80%"/>
<p>
## Quickstart
Below, we provide simple examples to show how to use our model with Transformers.
### Environment Setup
1. Download and navigate to the HuggingFace repository:
```
huggingface-cli download nvidia/omnivinci --local-dir ./omnivinci --local-dir-use-symlinks False
cd ./omnivinci
```
2. Install Python environment (based on NVILA codebase):
```
bash ./environment_setup.sh omnivinci
```
### 🤗 Transformers Usage
#### Video (with Audio) Inference Example
```python
from transformers import AutoProcessor, AutoModel, AutoConfig,AutoModelForCausalLM
import torch
import os
# default: Load the model on the available device(s)
model_path = "./"
video_path = "xxx.mp4"
generation_kwargs = {"max_new_tokens": 1024, "max_length": 99999999}
load_audio_in_video = True
num_video_frames = 128
audio_length = "max_3600"
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path,
trust_remote_code=True,
torch_dtype="torch.float16",
device_map="auto")
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
generation_config = model.default_generation_config
generation_config.update(**generation_kwargs)
model.config.load_audio_in_video = load_audio_in_video
processor.config.load_audio_in_video = load_audio_in_video
if num_video_frames > 0:
model.config.num_video_frames = num_video_frames
processor.config.num_video_frames = num_video_frames
if audio_length != -1:
model.config.audio_chunk_length = audio_length
processor.config.audio_chunk_length = audio_length
conversation = [{
"role": "user",
"content": [
{"type": "video", "video":video_path},
{"type": "text", "text": "Assess the video, followed by a detailed description of its video and audio contents."}
]
}]
text = processor.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)
inputs = processor([text])
output_ids = model.generate(
input_ids=inputs.input_ids,
media=getattr(inputs, 'media', None),
media_config=getattr(inputs, 'media_config', None),
generation_config=generation_config,
)
print(processor.tokenizer.batch_decode(output_ids, skip_special_tokens=True))
```
- **For audio and image inference examples, please refer to `example_mini_audio.py` and `example_mini_image.py`.**
## License / Terms of Use
The model is released under the [NVIDIA OneWay Noncommercial License](asset/NVIDIA_OneWay_Noncommercial_License.docx).
## Citation
Please consider to cite our paper and this framework, if they are helpful in your research.
```bibtex
@article{ye2025omnivinci,
title={OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM},
author={Ye, Hanrong and Yang, Chao-Han Huck and Goel, Arushi and Huang, Wei and Zhu, Ligeng and Su, Yuanhang and Lin, Sean and Cheng, An-Chieh and Wan, Zhen and Tian, Jinchuan and others},
journal={arXiv preprint arXiv:2510.15870},
year={2025}
}
``` |