🌋🌋 LLaVA-HR: High-Resolution Large Language-Vision Assistant 🌋🌋

✨Technical Report:

Feast Your Eyes: Mixture-of-Resolution Adaptation for Multimodal Large Language Models
Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, Rongrong Ji

This repository contains the implementation of LLaVA-HR, a strong and efficient MLLM powered by our mixture-of-resolution adaptation. The features of LLaVA-HR include:

• High Image Resolutions: LLaVA-HR supports up to 1536 x 1536 image resolutions, which boosts the performance of fine-grained vision-language tasks, such as TextVQA.
• Remarkable Efficiency: LLaVA-HR maintains the similar training costs with LLaVA-1.5, e.g., ~20 hours on 8 A100s. Its inference speed is also fast as existing low-resolution MLLMs ! Check out our paper.
• Strong Performance: LLaVA-HR outperforms existing MLLMs on multiple benchmarks, e.g., 82.6 on VQAv2. LLaVA-HR is comparable to LLaVA-NexT using the training data of LLaVA-1.5 ! Check out our model zoo.
• Fair Comparison: LLaVA-HR adopts the same training data and configurations with LLaVA-1.5, which means that the performance gains all come from our mixture-of-resolution adaptation. We hope that LLaVA-HR can be a strong baseline for the community.

📣 News

• [2024.04.16] We fix the evaluation bug for SQA and MMVet. Now, LLaVA-HR-X can achieve 40.3 score in MMVet! checking our model zoo.

• [2024.03.06] 🔥🔥🔥 We release LLaVA-HR, a high-resolution MLLM with strong performance and remarkable efficiency. LLaVA-HR greatly outperforms LLaVA-1.5 on multiple benchmarks, checking our model zoo.

Table of Contents

• Install
• Model Zoo
• Training
• Evaluation
• Demo

Install

1. Clone this repository and navigate to LLaVA-HR folder

git clone https://github.com/luogen1996/LLaVA-HR.git
cd LLaVA-HR

2. Install Package

conda create -n llava-hr python=3.10 -y
conda activate llava-hr
pip install --upgrade pip  # enable PEP 660 support
pip install -e .

3. Install additional packages for training cases

pip install ninja
pip install flash-attn --no-build-isolation

Model Zoo

Version	Size	Res	Checkpoint	VQAv2	GQA	VizWiz	TextVQA	OKVQA	OCRVQA	SQA	MME	POPE	SEED	MM-Vet
LLaVA-1.5	13B	336	liuhaotian/llava-v1.5-13b	80.0	63.3	53.6	61.3	-	-	71.6	1531.3	85.9	61.6	35.4
LLaVA-HR	7B	1024	favor123/llava-hr-7b-sft-1024	81.9	64.2	48.7	67.1	58.9	68.4	67.9	1554.9	87.6	64.2	31.5
LLaVA-HR-X	13B	1024	favor123/llava-hr-13b-x-sft-1024	82.6	65.2	56.6	70.9	61.5	69.0	69.7	1487.3	88.0	65.3	40.3

Training

Our training pipeline and datasets are directly borrowed from LLaVA-v1.5. The training consists of two stages:

• Low-resolution pretraining: train a projector on a subset of ∼558K image-text pairs to connect a frozen pretrained vision encoder and a frozen LLM.
• High-resolution instruction tuning: adopt our MR-Adaptation to accommodate high-resolution images, and fine-tune the whole MLLM with multimodal instruction data.

Training scripts

Stage-1: Low-resolution Pretraining

Please download the caption annotations blip_laion_cc_sbu_558k.json and images from here. Move the downloaded files to the /data/data folder. Then run the following command to start the training process:

bash scripts/v1_5/pretrain_llava_hr.sh

We recommend to directly use our pre-trained projector for better reproducing our results.

Version	Vision Encoder	Projection	Pretrain Data	Pretraining schedule	Download
LLaVA-HR-7b	CLIP-L & ConvNeXt-L	MLP-2x	LCS-558K	1e	projector
LLaVA-HR-X-13b	CLIP-L & ConvNeXt-XXL	MLP-2x	LCS-558K	1e	projector

Stage-2: High-resolution Instruction Tuning

Please download the annotation file of the mixed instruction tuning data llava_v1_5_mix665k.json, and download the images from constituting datasets:

• COCO: train2017
• GQA: images
• OCR-VQA: download script, save all files as .jpg
• TextVQA: train_val_images
• VisualGenome: part1, part2

After downloading all of them, organize the data as follows in ./playground/data:

├── coco
│   └── train2017
├── gqa
│   └── images
├── ocr_vqa
│   └── images
├── textvqa
│   └── train_images
└── vg
    ├── VG_100K
    └── VG_100K_2

Then, you can start the training process by the following script. If you use your custom dataset, you can refer to llava_v1_5_mix665k.json to format your data.

bash scripts/v1_5/train_eval_llava_hr.sh

Instruction tuning takes around 16 hours for LLaVA-HR-7B on 8x A100s (80G).

Evaluation

We follow LLaVA-v1.5 to conduct evaluations. you should download eval.zip and unzip it to ./playground/data/eval. Besides, we further implement the evaluation of coco-caption, refcoco, vizwiz,ocrvqa and okvqa. Please refer to Evaluation.md to prepare the data.

Then, your can run our evaluation script bash scripts/v1_5/eval.sh.

🤗 Demo

Gradio Web UI

Here are the steps to run the demo on your local devices.

Demo scripts

To launch a Gradio demo locally, please run the following commands one by one. If you plan to launch multiple model workers to compare between different checkpoints, you only need to launch the controller and the web server *ONCE*. #### Launch a controller ```Shell python -m llava.serve.controller --host 0.0.0.0 --port 10000 ```

Launch a gradio web server.

python -m llava.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload

You just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.

Launch a model worker

This is the actual worker that performs the inference on the GPU. Each worker is responsible for a single model specified in --model-path.

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path ./checkpoints/llava-hr-7b-sft-1024

Wait until the process finishes loading the model and you see "Uvicorn running on ...". Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.

You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the --controller the same, and modify the --port and --worker to a different port number for each worker.

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port <different from 40000, say 40001> --worker http://localhost:<change accordingly, i.e. 40001> --model-path <ckpt2>

If you are using an Apple device with an M1 or M2 chip, you can specify the mps device by using the --device flag: --device mps.

Launch a model worker (Multiple GPUs, when GPU VRAM <= 24GB)

If the VRAM of your GPU is less than 24GB (e.g., RTX 3090, RTX 4090, etc.), you may try running it with multiple GPUs. Our latest code base will automatically try to use multiple GPUs if you have more than one GPU. You can specify which GPUs to use with CUDA_VISIBLE_DEVICES. Below is an example of running with the first two GPUs.

CUDA_VISIBLE_DEVICES=0,1 python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path ./checkpoints/llava-hr-7b-sft-1024

CLI Inference

Here is the command for chatting with LLaVA-HR without the need of Gradio interface.

python -m llava.serve.cli \
    --model-path ./checkpoints/llava-hr-7b-sft-1024 \
    --image-file "./assets/example.jpg" 

👍 Acknowledgement

• LLaVA The codebase we built upon, and our baseline LLaVA-1.5 already has strong multimodal capabilities.

🔒 License

• The majority of this project is released under the Apache 2.0 license as found in the LICENSE file.
• The service is a research preview intended for non-commercial use only, subject to the model License of LLaMA and Terms of Use of the data generated by OpenAI. Please contact us if you find any potential violation.

✏️ Citation

If you find our paper and code useful in your research, please consider giving a star ⭐️ and citation 📝.

@article{luo2024feast,
  title={Feast Your Eyes: Mixture-of-Resolution Adaptation for Multimodal Large Language Models},
  author={Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, Rongrong Ji},
  journal={arXiv preprint arXiv:2403.03003},
  year={2024}
}

✨ Star History