---
{}
---

[![CODE](https://img.shields.io/badge/GitHub-Repository-<COLOR>)](https://github.com/bfshi/scaling_on_scales)

# When Do We Not Need Larger Vision Models?

## Model

This is a LLaVA-v1.5-7b model trained with [S<sup>2</sup>-Wrapper](https://github.com/bfshi/scaling_on_scales), a simple approach to enable any vision model to perceive high-resolution images. We use image resolutions of up to 1008x1008 for this model.

## Training

The training pipeline and dataset completely follow [LLaVA-v1.5](https://github.com/haotian-liu/LLaVA/tree/main). We use LoRA to fine-tune the model.

## Benchmarking

| Version | Size | Schedule | Checkpoint | VQAv2 | VizWiz | TextVQA | MMMU-val | MathVista | MM-Bench | SEED | MM-Vet |
|----------|----------|-----------|-----------|---|---|---|---|---|---|---|---|
| LLaVA-1.5 | 7B | full_ft-1e | [liuhaotian/llava-v1.5-7b](https://huggingface.co/liuhaotian/llava-v1.5-7b) | 78.5 | 50.0 | 58.2 | 36.2 | 25.2 | 64.3 | 65.7 | 31.1 |
| LLaVA-1.5 | 7B | lora-1e | [liuhaotian/llava-v1.5-7b-lora](https://huggingface.co/liuhaotian/llava-v1.5-7b-lora) | 79.1 | 47.8 | 58.2 | - | - | 66.1 | - | 30.2 |
| LLaVA-1.5-S2 | 7B | lora-1e | this model | **80.0** | **50.1** | **61.0** | **37.7** | **25.3** | **66.2** | **67.9** | **32.4** |



## License
Llama 2 is licensed under the LLAMA 2 Community License, 
Copyright (c) Meta Platforms, Inc. All Rights Reserved.