LLaMA-VID empowers existing frameworks to support hour-long videos and pushes their upper limit with an extra context token. We build this repo based on LLaVA.

• [24/07/04] 🔥 Our work has been accepted to ECCV 2024!
• [23/12/05] 🔥 We release the full training and evalution model, data, and scripts to support movie chating!
• [23/11/29] 🔥 LLaMA-VID is comming! We release the paper, code, data, models, and demo for LLaMA-VID!

• Demo
• Install
• Model
• Preparation
• Train
• Evaluation
• Examples
• Citation
• Acknowledgement
• License

We provide some selected examples in this section. More examples can be found in our project page. Feel free to try our online demo!

Please follow the instructions below to install the required packages.

1. Clone this repository

git clone https://github.com/dvlab-research/LLaMA-VID.git

2. Install Package

conda create -n llamavid python=3.10 -y
conda activate llamavid
cd LLaMA-VID
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

LLaMA-VID simply contains three parts: encoder and decoder are adopted to produce visual embedding and text-guided features, respectively; context token and content token are transformed with the tailored token generation strategy; instruction tuning is designed to unleash the potential of LLMs for image and video.

We provide all our fully finetuned models on Stage 1 and 2 data (Long Video + Stage 3) for LLaMA-VID:

Here are the pretrained weights (text decoder + context attention + projector) on Stage 1 data only:

We provide the processed image-based data for LLaMA-VID training. We organize the data in the format of LLaVA, please organize the training image-based data following this and evaluation image-based data following this. Please put the pretrained data, finetuned data, and eval data in LLaMA-VID-Pretrain, LLaMA-VID-Finetune, and LLaMA-VID-Eval subset following Structure.

For video-based dataset, please download the 2.5M subset from WebVid and ActivityNet dataset from official website or video-chatgpt. If you want to perform evaluation, please also download corresponding files from here. You can download MSVD-QA from here and MSRVTT-QA from here.

As for long video tuning, please download the long video data from MovieNet, shot detection results from here and our construced long video QA pairs from here. Place shot detection results under LLaMA-VID-Finetune/movienet/files before preprocessing.

For meta info, please download the following files and organize them as in Structure.

We recommend users to download the pretrained weights from the following link Vicuna-7b-v1.5, Vicuna-13b-v1.5, EVA-ViT-G, QFormer-7b, QFormer-13b and put them in model_zoo following Structure.

The folder structure should be organized as follows before training.

LLaMA-VID
├── llamavid
├── scripts
├── work_dirs
│   ├── llama-vid
│   │   ├── llama-vid-13b-full-336
│   │   ├── ...
├── model_zoo
│   ├── LLM
│   │   ├── vicuna
│   │   │   ├── 7B-V1.5
│   │   │   ├── 13B-V1.5
│   ├── LAVIS
│   │   ├── eva_vit_g.pth
│   │   ├── instruct_blip_vicuna7b_trimmed.pth
│   │   ├── instruct_blip_vicuna13b_trimmed.pth
├── data
│   ├── LLaMA-VID-Pretrain
│   │   ├── blip_laion_cc_sbu_558k.json
│   │   ├── llava_558k_with_webvid.json
│   │   ├── images
│   │   ├── videos
│   ├── LLaMA-VID-Finetune
│   │   ├── llava_v1_5_mix665k.json
│   │   ├── llava_v1_5_mix665k_maxround_6_total_921k.json
│   │   ├── llava_v1_5_mix665k_maxround_12_total_714k.json
│   │   ├── llava_v1_5_mix665k_with_video_chatgpt.json
│   │   ├── llava_v1_5_mix665k_with_video_chatgpt_maxtime_5min.json
│   │   ├── long_videoqa.json
│   │   ├── movienet
│   │   ├── activitynet
│   │   ├── coco
│   │   ├── gqa
│   │   ├── ocr_vqa
│   │   ├── textvqa
│   │   ├── vg
│   ├── LLaMA-VID-Eval
│   │   ├── gqa
│   │   ├── ...

LLaMA-VID training consists of three stages: (1) feature alignment stage: bridge the vision and language tokens; (2) instruction tuning stage: teach the model to follow multimodal instructions; (3) long video tuning stage: extend the position embedding and teach the model to follow hour-long video instructions.

LLaMA-VID is trained on 8 A100 GPUs with 80GB memory. To train on fewer GPUs, you can reduce the per_device_train_batch_size and increase the gradient_accumulation_steps accordingly. Always keep the global batch size the same: per_device_train_batch_size x gradient_accumulation_steps x num_gpus.

Please make sure you download and organize the data following Preparation before training.

If you only want to train and finetune LLaMA-VID on image-based data, please run the following command for Vicuna-7B with image size 336:

bash scripts/image_only/train/stage_1_2_full_v7b_336.sh

or for Vicuna-13B with image size 336:

bash scripts/image_only/train/stage_1_2_full_v13b_336.sh

You can also try that with a smaller image size 224 and less visual tokens:

bash scripts/image_only/train/stage_1_2_full_v7b_224_grid_4.sh

Please find more training scripts in scripts/image_only/train.

If you are interested in training and finetuning LLaMA-VID on short video-based data, please run the following command for Vicuna-7B with image size 224:

bash scripts/video/train/stage_1_2_full_v7b_224_fps_1.sh

or for Vicuna-13B with image size 224:

bash scripts/video/train/stage_1_2_full_v13b_224_fps_1.sh

Please find more training scripts in scripts/video/train.

We provide dataset and scripts for long video-based training. Please download the long video-based data following Preparation and organize them as in Structure. In the training stage, we first extract all the frames from the long video and save the visual features local for efficient training.

python scripts/extra_tool/extract_movienet_features.py \
    --video_dir <path to movienet video> \
    --files_dir <path to movienet files> \ # files in downladed MovieNet.tar.gz
    --feat_dir <path to output features>

And run the following command for Vicuna-7B with image size 224:

bash scripts/video/train/stage_3_full_v7b_224_longvid.sh

We perform evaluation on both image-based and video-based benchmarks. Please download the evaluation data following Preparation and organize them as in Structure.

If you want to evaluate the model on image-based benchmarks, please use the scripts in scripts/image_only/eval. For example, run the following command for GQA evaluation:

bash scripts/image_only/eval/gqa.sh

Please find more evaluation scripts in scripts/image_only/eval.

If you want to evaluate the model on video-based benchmarks, please use the scripts in scripts/video/eval. For example, run the following command for MSVD-QA evaluation:

bash scripts/video/eval/msvd_eval.sh

Please find more evaluation scripts in scripts/video/eval.

Chat with images and videos using LLaMA-VID without the need of Gradio interface. It also supports multiple GPUs, 4-bit and 8-bit quantized inference. With 4-bit quantization. Please try this for image or video inference:

python -m llamavid.serve.cli \
    --model-path work_dirs/llama-vid/llama-vid-7b-full-336 \
    --image-file <path to your image>

or try this for video inference:

python -m llamavid.serve.cli \
    --model-path work_dirs/llama-vid/llama-vid-7b-full-224-video-fps-1 \
    --image-file <path to your video> \
    --temperature 0.5

You can also try 4bit or 8bit for efficient inference

python -m llamavid.serve.cli \
    --model-path work_dirs/llama-vid/llama-vid-7b-full-224-video-fps-1 \
    --image-file <path to your video>
    --temperature 0.5 \
    --load-4bit

For long video, if you want to inference on videos in movienet, please first process the video data and subtitles like this:

python scripts/extra_tool/extract_movienet_features.py \
    --video_dir <path to movienet video> \
    --files_dir <path to movienet files> \ # files in downladed MovieNet.tar.gz
    --feat_dir <path to output features>

If you want to inference with your customized video, please first process the video data and subtitles like this:

python scripts/extra_tool/extract_video_features_subtitles.py \
    --video_file <path to customized video> \
    --feat_dir <path to output features>

Then, please try this for long video inference:

python llamavid/serve/run_llamavid_movie.py \
    --model-path work_dirs/llama-vid/llama-vid-7b-full-224-long-video \
    --video-file <path to your processed video file> \
    --load-4bit

Here, we adopt the Gradio UI similar to that in LLaVA to provide a user-friendly interface for LLaMA-VID. 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.

python -m llamavid.serve.controller --host 0.0.0.0 --port 10000

python -m llamavid.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.

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 llamavid.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path work_dirs/llama-vid/llama-vid-vicuna-7b-short

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 models in the same Gradio interface. For example, short video model here. Please keep the --controller the same, and modify the --port and --worker to a different port number for each worker.

python -m llamavid.serve.model_worker_short --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 work_dirs/llama-vid/llama-vid-7b-full-224-video-fps-1

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.

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 llamavid.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path work_dirs/llama-vid/llama-vid-7b-full-224-long-video

You can launch the model worker with quantized bits (4-bit, 8-bit), which allows you to run the inference with reduced GPU memory footprint. Note that inference with quantized bits may not be as accurate as the full-precision model. Simply append --load-4bit or --load-8bit to the model worker command that you are executing. Below is an example of running with 4-bit quantization.

python -m llamavid.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path work_dirs/llama-vid/llama-vid-7b-full-224-long-video --load-4bit

We provide some examples in this section. More examples can be found in our project page.

If you find this repo useful for your research, please consider citing the paper

@inproceedings{li2024llamavid,
  title={LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models},
  author={Li, Yanwei and Wang, Chengyao and Jia, Jiaya},
  journal={European Conference on Computer Vision},
  year={2024}
}

We would like to thank the following repos for their great work:

• This work is built upon the LLaVA.
• This work utilizes LLMs from Vicuna.
• This work utilizes pretrained weights from InstructBLIP.
• We perform video-based evaluation from Video-ChatGPT.

The data and checkpoint is intended and licensed for research use only. They are also restricted to uses that follow the license agreement of LLaVA, LLaMA, Vicuna and GPT-4. The dataset is CC BY NC 4.0 (allowing only non-commercial use) and models trained using the dataset should not be used outside of research purposes.