---
license: mit
---
# 🔥 SPHINX: A Mixer of Tasks, Domains, and Embeddings

Official implementation of ['SPHINX: A Mixer of Tasks, Domains, and Embeddings Advances Multi-modal Large Language Models'](https://github.com/Alpha-VLLM/LLaMA2-Accessory/tree/main/SPHINX).

Try out our [web demo 🚀](http://imagebind-llm.opengvlab.com/) here!

<p align="left">
   Github link: <a href="https://huggingface.co/Alpha-VLLM/SPHINX" target="_blank">Github</a> • 👋 join our <a href="https://github.com/Alpha-VLLM/LLaMA2-Accessory/blob/main/docs/wechat.md" target="_blank">WeChat</a>
</p>


## Introduction

We present SPHINX, a versatile multi-modal large language model (MLLM) with a mixer of training tasks, data domains, and visual embeddings. 

- **Task Mix.** For all-purpose capabilities, we mix a variety of vision-language tasks for mutual improvement: VQA, REC, REG, OCR, etc.

- **Embedding Mix.** We capture robust visual representations by fusing distinct visual architectures, pre-training, and granularity.

- **Domain Mix.** For data from real-world and synthetic domains, we mix the weights of two domain-specific models for complementarity.

<p align="left">                                                                                                                                          
  <img src="figs/pipeline1.png"/ width="100%"> <br>
</p>

On top of SPHINX, we propose to further mixvisual scales and sub-images for better capture fine-grained semantics on high-resolution images.
<p align="left">                                                                                                                                          
  <img src="figs/pipeline2.png"/ width="100%"> <br>
</p>


### Installation
SPHINX is built upon LLaMA2-Accessory, please follow the instructions [here](https://llama2-accessory.readthedocs.io/en/latest/install.html) for environment setup.



## Inference
This section provides a step-by-step guide for hosting a local SPHINX demo. If you're already familiar with the LLAMA2-Accessory toolkit, note that hosting a SPHINX demo follows the same pipeline as hosting demos for the other models supported by LLAMA2-Accessory.


### Weights
We provide the beta-version checkpoints on [HuggingFace🤗](https://huggingface.co/Alpha-VLLM/LLaMA2-Accessory/tree/main/finetune/mm/sphinx-sft). Please download them to your own machine. The file structure should appear as follows:
```
ckpt_path/
├── consolidated.00-of-02.model.pth
└── consolidated.01-of-02.model.pth
```

### Host Local Demo
Execute the following command for demo hosting:
``` bash
cd LLaMA2-Accessory/accessory
python demos/multi_turn_mm.py --n_gpus=2 \
--tokenizer_path=/path/to/tokenizer.model --llama_type=llama_ens \
--pretrained_path ckpt_path/
```
Explanation of each argument:

+ `--n_gpus`: Number of gpus to use. Utilizing more GPUs will alleviate memory usage on each GPU through model parallelism. Currently, this argument should be set to either 1 or 2, as support for *consolidated ckpt num < gpu num* is not yet available.
+ `--tokenizer_path`: Path to the official LLaMA2 tokenizer. Note that the tokenizer file is the same for both LLaMA and LLaMA2. You may download it from [here](https://huggingface.co/Alpha-VLLM/LLaMA2-Accessory/blob/main/config/tokenizer.model).
+ `--llama_type`: The model architecture of SPHINX is defined in [accessory/model/LLM/llama_ens.py](../accessory/model/LLM/llama_ens.py),  and specifying `--llama_type=llama_ens ` tells the demo program to use this architecture.
+ `--pretrained_path`: The path to pre-trained checkpoint.


## Result

We provide a comprehensive evaluation of SPHINX and showcase results across multiple benchmarks. 

Our evaluation encompasses both **quantitative metrics** and **qualitative assessments**, providing a holistic understanding of our VLM model's performance.

**Evaluation Prompt Design**
<p align="left">                                                                                                                                          
  <img src="figs/table1.png"/ width="100%"> <br>
</p>

* In evaluation, we prioritize aligning with each benchmark's desired output format. 
* We employ distinct prompts tailored to benchmarks that necessitate long answers, short answers, and multiple-choice responses.
* For tasks involving visual grounding, we directly utilize the prompts during training to enhance the model's performance on these particular challenges.

**Benchmarks on Multimodal Large Language Models**
<p align="left">                                                                                                                                          
  <img src="figs/table2.png"/ width="100%"> <br>
</p

* We test our model on recently proposed MLLM benchmarks which is based on VQA to comprehensive evaluation of the model's characteristic such as MME, Seedbench, POPE, LLaVA-Bench (In-the-Wild), MM-Vet, MathVista, MMbench, CCbench. 
* The Long-SPHINX achieve new stat of arts result on 5 out of 9 benchmarks

**Visual Question Answering**
<p align="left">                                                                                                                                          
  <img src="figs/table3.png"/ width="100%"> <br>
</p>

* We evaluate general VQA benchmarks, such as VQAV2, OKVQA, GQA, vizwiz, scienceQA, visual spatial reasoning (VSR), IconQA. 
* Additionally, we conduct experiments on Text-oriented VQA such as TextVQA,OCR-VQA.
* Long-Sphinx achieve comparative results across all benchmarks. We observe that Long-Sphinx outperforms Sphinx in VQA datasets that demand fine-grained visual information, showcasing the effectiveness of our visual mixed-up approach for achieving high resolution without relying on a visual encoder trained specifically on high-resolution images.

**Visual Grounding**
<p align="left">                                                                                                                                          
  <img src="figs/table4.png"/ width="100%"> <br>
</p>

* The SPHINX model and baseline models on REC benchmarks results on table4.
* SPHINX exhibits robust performance in visual grounding tasks such as RefCOCO, RefCOCO+, and RefCOCOg, **surpassing other vision-language generalist models**.
* Notably, SPHINX outperforms specialist models G-DINO-L by **more than 1.54%** in accuracy across all tasks within RefCOCO/RefCOCO+/RefCOCOg.