README.md

---
license: cc
tags:
- multimodal
- conversational
- GGUF
- Image-Text-to-Text
---
# Omnivision

## Introduction

Omnivision is a compact, sub-billion (968M) multimodal model for processing both visual and text inputs, optimized for edge devices. Built on LLaVA's architecture, it features:

- **9x Token Reduction**: Reduces image tokens from 729 to 81, cutting latency and computational cost.
- **Minimal-Edit DPO**: Enhances response quality with minimal edits, preserving core model behavior.
  
**Quick Links:**
1. Interactive Demo in our [Hugging Face Space](https://huggingface.co/spaces/NexaAIDev/omnivlm-dpo-demo).
2. [Quickstart for local setup](#how-to-use-on-device)
3. Learn more in our [Blogs](https://nexa.ai)

**Feedback:** Send questions or comments about the model in our [Discord](https://discord.gg/nexa-ai)

## Intended Use Cases
Omnivision is designed for **Visual Question Answering** (answering questions about images) and **Image Captioning** (describing scenes in photos), making it ideal for on-device applications.

**Example Demo:**
Omni-Vision generated captions for a 1046×1568 pixel poster | **Processing time: <2s** | Device: MacBook M4 Pro

<img src="https://cdn-uploads.huggingface.co/production/uploads/6618e0424dbef6bd3c72f89a/PTG3_n_p7_atBHCwRLOEE.png" alt="Example" style="width:700px;"/>


## Benchmarks

Below we demonstrate a figure to show how Omnivision performs against nanollava. In all the tasks, omnivision outperforms the previous world's smallest vision-language model.

<img src="https://cdn-uploads.huggingface.co/production/uploads/6618e0424dbef6bd3c72f89a/KsN-gTFM5MfJA5E3aDRJI.png" alt="Benchmark Radar Chart" style="width:500px;"/>

We have conducted a series of experiments on benchmark datasets, including MM-VET, ChartQA, MMMU, ScienceQA, POPE to evaluate the performance of omnivision.

| Benchmark         | Nexa AI Omni-Vision | nanoLLAVA | Qwen2-VL-2B |
|-------------------|----------------------|-----------|-------------|
| MM-VET            | 27.5                | 23.9      | 49.5        |
| ChartQA (Test)    | 59.2                | NA        | 73.5        |
| MMMU (Test)       | 41.8                | 28.6      | 41.1        |
| MMMU (Eval)       | 39.9                | 30.4      | 41.1        |
| ScienceQA (Eval)  | 62.2                | 59.0      | NA          |
| ScienceQA (Test)  | 64.5                | 59.0      | NA          |
| POPE              | 89.4                | 84.1      | NA          |


## How to Use On Device
In the following, we demonstrate how to run omnivision locally on your device.

**Step 1: Install Nexa-SDK (local on-device inference framework)**

[Install Nexa-SDK](https://github.com/NexaAI/nexa-sdk?tab=readme-ov-file#install-option-1-executable-installer)

> Nexa-SDK is a open-sourced, local on-device inference framework, supporting text generation, image generation, vision-language models (VLM), audio-language models, speech-to-text (ASR), and text-to-speech (TTS) capabilities. Installable via Python Package or Executable Installer.

**Step 2: Then run the following code in your terminal**

```bash
nexa run omnivision 
```

## Model Architecture ##
Omni-Vision's architecture consists of three key components:

- Base Language Model: Qwen2.5-0.5B-Instruct functions as the base model to process text inputs
- Vision Encoder: SigLIP-400M operates at 384 resolution with 14×14 patch size to generate image embeddings
- Projection Layer: Multi-Layer Perceptron (MLP) aligns the vision encoder's embeddings with the language model's token space

The vision encoder first transforms input images into embeddings, which are then processed by the projection layer to match the token space of Qwen2.5-0.5B-Instruct, enabling end-to-end visual-language understanding.

## Training

We developed Omni-Vision through a three-stage training pipeline:

**Pretraining:**
The initial stage focuses on establishing basic visual-linguistic alignments using image-caption pairs, during which only the projection layer parameters are unfrozen to learn these fundamental relationships.

**Supervised Fine-tuning (SFT):**
We enhance the model's contextual understanding using image-based question-answering datasets. This stage involves training on structured chat histories that incorporate images for the model to generate more contextually appropriate responses.

**Direct Preference Optimization (DPO):**
The final stage implements DPO by first generating responses to images using the base model. A teacher model then produces minimally edited corrections while maintaining high semantic similarity with the original responses, focusing specifically on accuracy-critical elements. These original and corrected outputs form chosen-rejected pairs. The fine-tuning targeted at essential model output improvements without altering the model's core response characteristics

## What's next?
We are continually improving Omnivision for better on-device performance. Stay tuned.

### Learn more in our blogs
[Blogs](https://nexa.ai)
### Join Discord Community
[Discord](https://discord.gg/nexa-ai)