Skip to content

TencentQQGYLab/StableLLAVA

 
 

Repository files navigation

Official repo for StableLLaVA

StableLLaVA: Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data

Yanda Li, Chi Zhang, Gang Yu, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, Yunchao Wei

[Arxiv] [Project Page]

This repository offers a collection of AI-generated datasets specifically tailored for visual instruction tuning.

Install

conda create -n stablellava python=3.10 -y
conda activate stablellava
pip install --upgrade pip 
pip install -e .

Evaluation

MMBench

  1. Download mmbench dev/test set, then put it under ./playground/data/eval/mmbench

  2. Set model_path and model_base in scripts/v1_5/eval/mmbench.sh

    For model_path, you can download from google drive

    For model_base, we adopt vicuna-13b-v1.5, and you can download from huggingface

    Test with single-gpu

    CUDA_VISIBLE_DEVICES=0 bash scripts/v1.5/eval/mmbench.sh

  3. Submit the results under ./playground/data/eval/mmbench/answers_upload/ to MMBench

Abstract

The remarkable multimodal capabilities demonstrated by OpenAI's GPT-4 have sparked significant interest in the development of multimodal Large Language Models (LLMs). A primary research objective of such models is to align visual and textual modalities effectively while comprehending human instructions. Current methodologies often rely on annotations derived from benchmark datasets to construct image-dialogue datasets for training purposes, akin to instruction tuning in LLMs. However, these datasets often exhibit domain bias, potentially constraining the generative capabilities of the models. In an effort to mitigate these limitations, we propose a novel data collection methodology that synchronously synthesizes images and dialogues for visual instruction tuning. This approach harnesses the power of generative models, marrying the abilities of ChatGPT and text-to-image generative models to yield a diverse and controllable dataset with varied image content. Additionally, datasets can be arbitrarily scaled. This not only provides greater flexibility compared to existing methodologies but also significantly enhances several model capabilities. Our research includes comprehensive experiments conducted on various datasets. The results emphasize substantial enhancements in more than ten commonly assessed capabilities. Additionally, our model achieves state-of-the-art results across multiple widely recognized multimodal benchmarks.

Pipeline

The prompt-dialogue of varies abilities are saved in dataset.

The synthesized prompt-dialogue datasets of various abilities are saved in dataset. Please follow the steps below to generate datasets with LLaVA format.

  1. Use SD-XL to generate images as training images. It will take ~13s to generate one image on V100.
python stable_diffusion.py --prompt_path dataset/animal.json --save_path train_set/animal/
  1. Use data_to_llava to convert dataset format for LLaVA model training.
python data_to_llava.py --image_path train_set/ --prompt_path dataset/ --save_path train_ano/

Results

Qualitative Results

Quantitative Results

  1. For training with the generated datasets, you can use LLaVA official code LLaVA

TO-DO LIST

  • Update more benchmark evaluation
  • Update multi-image evaluation
  • Keep incorporating more capabilities

Releases

No releases published

Packages

No packages published

Languages

  • Python 94.4%
  • JavaScript 2.8%
  • HTML 2.1%
  • Other 0.7%