github.com/JavisVerse/JavisGPT @main

JavisGPT: A Unified Multi-modal LLM for Sounding-Video Comprehension and Generation

[HomePage] [HF Paper] [ArXiv Paper] [Model] [Dataset]

TL;DR

We introduce JavisGPT, a multimodal LLM that can understand audiovisual inputs and simultaneously generate synchronized sounding videos in a unified model. We also curate the JavisInst-Omni dataset to facilitate instruction-tuning for comprehension and generation on sounding videos.

📰 News

• [2026.2.26] 🔥🔥 We release the upgraded JavisGPT-v1.0-7B-Instruct checkpoint at huggingface, which is empowered by JavisDiT-v1.0-jav to achieve better audio-video generation.
• [2026.1.1] 🚀 We release the full training and evaluation scripts to support future research in the community. Have fun with them!
• [2025.12.30] 🚀 We release the training dataset of JavisInst-Omni to support multimodal instruction tuning on sounding video comprehension and generation tasks, as well as MM-PreTrain and AV-FineTune datasets to enable preliminary multimodal alignment for LLMs. The JavisUnd-Eval dataset is also released to set a standard for audio-video understanding evaluation for MLLMs.
• [2025.12.26] 🔥 We release the code of JavisGPT, with the preview JavisGPT-v0.1-7B-Instruct checkpoint at huggingface. Feel free to play with it!

Code

We assume the data structure as:

/path/to/user/root
|-- projects
|   |   |-- JavisDiT  # downstream JAV-DiT
|   |   └-- JavisGPT  # workspace of this project
|-- weights
|   |-- pretrained
|   |   |-- dit   # pretrained weights for JavisDiT
|   |   |   |-- OpenSora-VAE-v1.2
|   |   |   |-- audioldm2
|   |   |   |-- t5-v1_1-xxl
|   |   |   |-- Wan2.1-T2V-1.3B
|   |   |   |-- imagebind_huge.pth
|   |   |-- mllm  # pretrained weights for JavisGPT
|   |   |   |-- BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt
|   |   |   └-- Qwen2.5-VL-7B-Instruct
|   |-- JavisVerse
|   |   |-- JavisDiT-v0.1-prior
|   |   |-- JavisDiT-v0.1-jav-240p4s
|   |   |-- JavisGPT-v0.1-7B-Instruct
|   |   |-- JavisDiT-v1.0-jav
|   |   └-- JavisGPT-v1.0-7B-Instruct
|-- datasets
|   |-- JavisGPT
|   |   |-- train
|   |   |   |-- MM-PreTrain
|   |   |   |-- AV-FineTune
|   |   |   └-- JavisInst-Omni
|   |   |-- eval
|   |   |   |-- JavisUnd-Eval
|   |   |   └-- JavisBench

Installation

Install the necessary packages:

conda create -n javisgpt python=3.10 -y
conda activate javisgpt
pip install --upgrade pip  # Enable PEP 660 support.
pip install flash-attn==2.7.4.post1 --no-build-isolation
pip install -v -e ".[train]"
cp assets/src/dynamic_modules_utils.py /path/to/python3.10/site-packages/diffusers/utils/
conda install "ffmpeg<7" -c conda-forge -y  # install ffpmeg

Install JavisDiT dependencies:

cd ..
git clone https://github.com/JavisVerse/JavisDiT.git
cd JavisDiT
pip install -v -e . --no-deps
cd ../JavisGPT

# # make soft links if necessary
# ln -s ../JavisDiT/javisdit javisdit

Inference

1. Prepare Pretrained Weights

First, download BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt from here and Qwen2.5-VL-7B-Instruct, and put (or link) them into ../../weights/pretrained/mllm.

hf download Qwen/Qwen2.5-VL-7B-Instruct --local-dir ../../weights/pretrained/mllm/Qwen2.5-VL-7B-Instruct

Then, download our JavisGPT-v0.1-7B-Instruct and JavisGPT-v1.0-7B-Instruct and put them into ../../weights/JavisVerse, e.g.,

# (Optional) for JavisGPT-v0.1-7B-Instruct
hf download JavisVerse/JavisGPT-v0.1-7B-Instruct --local-dir ../../weights/JavisVerse/JavisGPT-v0.1-7B-Instruct

# for JavisGPT-v1.0-7B-Instruct
hf download JavisVerse/JavisGPT-v1.0-7B-Instruct --local-dir ../../weights/JavisVerse/JavisGPT-v1.0-7B-Instruct

Finally, download necessary checkpoints of the downstream JAVG model (JavisDiT) and put them into ../../weights/pretrained/dit or ../../weights/JavisVerse, according to path definition in ./interface/config/*.py coordinately.

# (Optional) for JavisGPT-v0.1-7B-Instruct
# hf download hpcai-tech/OpenSora-VAE-v1.2 --local-dir ../../weights/pretrained/dit/OpenSora-VAE-v1.2
# hf download cvssp/audioldm2 --local-dir ../../weights/pretrained/dit/audioldm2
# hf download DeepFloyd/t5-v1_1-xxl --local-dir ../../weights/pretrained/dit/t5-v1_1-xxl
# hf download JavisVerse/JavisDiT-v0.1-prior --local-dir ../../weights/JavisVerse/JavisDiT-v0.1-prior
# hf download JavisVerse/JavisDiT-v0.1-jav-240p4s --local-dir ../../weights/JavisVerse/JavisDiT-v0.1-jav-240p4s

# for JavisGPT-v1.0-7B-Instruct
hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir ../../weights/pretrained/dit/Wan2.1-T2V-1.3B
hf download cvssp/audioldm2 --local-dir ../../weights/pretrained/dit/audioldm2
hf download JavisVerse/JavisDiT-v1.0-jav --local-dir ../../weights/JavisVerse/JavisDiT-v1.0-jav

2. Run Target Inference

• Standalone Audio/Visual Comprehension

Use the following commands to evaluate the preserved single-modality understanding capability.

For audio comprehension:

AUDIO_PATH="assets/demos/audio/Creaking_pier.wav"
PROMPT="Is the sound caused by pressure from/against wood?"
JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} AUDIO_PATH=${AUDIO_PATH} PROMPT=${PROMPT} \
bash scripts/demo/demo_audio_visual.sh

For video comprehension:

VIDEO_PATH="assets/demos/video/ZS9XR.mp4"
PROMPT="What happened after the person took the box? A. Ate the medicine. B. Tidied up the blanket. C. Put down the cup/glass/bottle. D. Open the computer."
JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} VIDEO_PATH=${VIDEO_PATH} PROMPT=${PROMPT} \
bash scripts/demo/demo_audio_visual.sh

• Joint Audio-Video Comprehension

Use the following command to evaluate the joint audio-video comprehension capability.

VIDEO_PATH="assets/demos/audio_video/00002617.mp4"
PROMPT="How many instruments in the room did not sound from beginning to end? Answer the question using a single word."
USE_AUDIO_IN_VIDEO=True
JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} VIDEO_PATH=${VIDEO_PATH} PROMPT=${PROMPT} USE_AUDIO_IN_VIDEO=${USE_AUDIO_IN_VIDEO} \
bash scripts/demo/demo_audio_visual.sh

• Joint Audio-Video Generation

Use the following command to evaluate the sounding video generation capability.

PROMPT="Build a video, ensuring the content is echoed by complementary scenes: A beautiful waterfall cascades down a steep cliff into a clear pool below. Sunlight filters through the surrounding trees, creating shimmering reflections on the falling water. The scene is calm and natural, with continuous flowing water and gentle mist rising from the base. The sound consists of steady rushing water, soft splashes, and faint ambient forest noise."
AV_GENERATE=True
SAVE_PREFIX="./results/avgen/demo"
JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} AV_GENERATE=${AV_GENERATE} PROMPT=${PROMPT} SAVE_PREFIX=${SAVE_PREFIX} \
bash scripts/demo/demo_audio_visual.sh

The generated sample will be saved at ${SAVE_PREFIX}.mp4, e.g., ./results/avgen/demo.mp4.

Training

1. Prepare Training Datasets

Download the corresponding MM-PreTrain, AV-FineTune, and JavisInst-Omni from HuggingFace, and put (or link) them into ../../datasets/JavisGPT/train/ coordinately.

hf download --repo-type dataset JavisVerse/MM-PreTrain --local-dir ../../datasets/JavisGPT/train/MM-PreTrain
hf download --repo-type dataset JavisVerse/AV-FineTune --local-dir ../../datasets/JavisGPT/train/AV-FineTune
hf download --repo-type dataset JavisVerse/JavisInst-Omni --local-dir ../../datasets/JavisGPT/train/JavisInst-Omni

Then, go to the downloaded datasets and run the unzip.py to extract source audio/video data:

cd ../../datasets/JavisGPT/train/MM-PreTrain # MM-PreTrain/AV-FineTune/JavisInst-Omni
python unzip.py --purge

However, we cannot release the source data of TAVGBench due to policy issues. Instead, the video_ids (formatted with {youtube_id}_{start_time}_{end_time}) are provided in video_ids.txt at AV-FineTune and JavisInst-Omni, and users can refer to TAVGBench to download raw videos.

2. Conduct Progressive Training Procedure

• Stage-I Multi-Modal Pre-Training

This stage aims to enquip the backbone Qwen2.5-VL-Instruct model with preliminary audio comprehension and audio-video generation capabilities, includes two distinct tasks: (1) audio comprehension pretraining and (2) audio–video generation pretraining. We adopt separate training for data efficiency, since the two tasks do not share gradient interactions during optimization.

For audio comprehension pretraining, run the following command, and the trained audio_proj.bin will be saved at ./runs/javisgpt_stage1_mm_pretrain_audio_align/:

bash ./scripts/train/train_audio_align.sh

For audio-video generation pretraining, run the following command, and the trained avgen_proj.bin will be saved at ./runs/javisgpt_stage1_mm_pretrain_avgen_align/:

JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} bash ./scripts/train/train_audio_video_gen_align.sh

• Stage-II Audio-Visual Fine-Tuning

This stage aims to enhance the understanding and generation of sounding videos, which can be integrated into a single task. Run the following command, and the trained mm_proj_all.bin with LoRA weights will be saved at ./runs/javisgpt_stage2_av_finetune/:

JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} bash ./scripts/train/stage2_av_ft.sh

• Stage-III Multi-Modal Insturction-Tuning

This stage aims to elicit the multimodal instruction-following ability for audio/video/audio-video comprehension and joint audio-video generation tasks. Run the following command, and the trained mm_proj_all.bin with LoRA weights will be saved at ./runs/javisgpt_stage3_mm_insttune/:

JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} bash ./scripts/train/stage3_mm_it.sh

Only the checkpoints trained in this stage are utilized to build the final JavisGPT model.

Evaluation

1. Prepare Evaluation Datasets

• Audio/Video/Audio-Video Comprehension

Download and extract the preprocessed JavisUnd-Eval dataset (including 8 widely-used subsets) to ../../datasets/JavisGPT/eval/:

hf download --repo-type dataset JavisVerse/JavisUnd-Eval --local-dir ../../datasets/JavisGPT/eval/JavisUnd-Eval

cd ../../datasets/JavisGPT/eval/JavisUnd-Eval
python unzip.py --purge
cd -

• Joint Audio-Video Generation

Download the JavisBench dataset to ../../datasets/JavisGPT/eval/:

hf download --repo-type dataset JavisVerse/JavisBench --local-dir ../../datasets/JavisGPT/eval/JavisBench

2. Run Target Inference

• Audio/Video/Audio-Video Comprehension

Run the following script to automatically collect the responses from audio (ClothoAQA, TUT2017), video (ActivityNet, Perception, MVBench), and audio-video (AVQA, MusicAVQA, AVSD) datasets:

JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} bash ./scripts/eval/infer_av_und.sh

Model's responses will be saved at ./results/av_und/.

• Joint Audio-Video Generation

Run the following script to automatically collect the joint audio-video generation results from JavisBench-mini:

JAV_VERSION="v1.0"

JAV_VERSION=${JAV_VERSION} bash ./scripts/eval/infer_av_gen.sh

Model's responses will be saved at ./results/av_gen/.

3. Run Target Evaluation

• Audio/Video/Audio-Video Comprehension

If you use local models as the LLM judge, such as Qwen series, install the vllm dependency and run the following commands:

pip install vllm==0.7.3

export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
export VLLM_WORKER_MULTIPROC_METHOD="spawn"

judge_model_name_or_path="Qwen/Qwen2.5-72B-Instruct"
res_dir="./results/av_und/"

python javisgpt/eval/eval_und.py \
    --res_dir "${res_dir}" \
    --judge_model_name_or_path "${judge_model_name_or_path}"

Otherwise, if you use OpenAI API as the LLM judge, run:

```bash api_key="YOU_API_KEY" judge_model_name_or_path="OpenAI/gpt-4o-mini" res_dir="./results/av_und/"

python javisgpt/eval/eval_und.py \ --res_dir "${res_dir}" \ --judge_model_name_or_path "${judge_model_name_or_path}"