Oryx-mllm/Oryx

[ICLR 2025] MLLM for On-Demand Spatial-Temporal Understanding at Arbitrary Resolution

Python

329

47 commits

updated Jul 4, 2025

oryx-mllm.github.io

See the code

README

Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution

Zuyan Liu^*,1,2 Yuhao Dong^*,2,3 Ziwei Liu³ Winston Hu² Jiwen Lu^1,✉ Yongming Rao^2,1,✉

¹Tsinghua University ²Tencent ³S-Lab, NTU

^* Equal Contribution^✉ Corresponding Author

Project Page:

arXiv Paper:

Demo by Gradio:

Model Checkpoints:

Oryx SFT Data:

📢 News

[23/1/2025] 🎉Oryx has been accepted to ICLR 2025!
[23/10/2024] 🎉Oryx-1.5 Series is released! Oryx-1.5 includes 7B and 32B variants. We achieve stronger performance on all the benchmarks! Check our results at VideoMME Leaderboard and the updated arXiv paper.
[30/09/2024] 📊Oryx video data for SFT is released!
[26/09/2024] 🎨Try out our online demo with Oryx-7B for image/video understanding!
[24/09/2024] 🚀 Oryx-34B is now available at VideoMME Leaderboard, Oryx-34B achieves best accuracy among <40B MLLMs.
[23/09/2024] 🔥 Oryx ranks 1st on MLVU, surpassing GPT-4o. Stay tuned for more results!
[20/09/2024] 🔥 🚀Introducing Oryx! The Oryx models (7B/34B) support on-demand visual perception, achieve new state-of-the-art performance across image, video and 3D benchmarks, even surpassing advanced commercial models on some benchmarks.
- [Paper]: Detailed introduction of on-demand visual perception, including native resolution perception and dynamic compressor!
- [Checkpoints]: Try our advanced model on your own.
- [Scripts]: Start training models with customized data.

🐐 Introducing Oryx

Oryx is a unified multimodal architecture for the spatial-temporal understanding of images, videos, and multi-view 3D scenes. Oryx offers an on-demand solution to seamlessly and efficiently process visual inputs with arbitrary spatial sizes and temporal lengths. Our model achieve strong capabilities in image, video, and 3D multimodal understanding simultaneously.

Main idea of On-Demand Multimodal Understanding

Overview of Oryx Architecture

✅ TODO List

Release all the model weights.
Release OryxViT model.
Demo code for generation.
All the training and inference code.
Evaluation code for image, video and 3D multi-modal benchmark.
Oryx SFT Data.
Oryx Gradio chatbox.
Enhanced Oryx model with latest LLM base models and better SFT data.
Introducing our explorations for OryxViT.

📃 Main Results

Results on General Temporal Understanding

Results on Long-Form Temporal Understanding

Results on Image Understanding

Results on 3D Understanding

Model Zoo

We provide our checkpoints at Huggingface

Model	Link	Size	Visual Encoder	LLM-Type	Intermediate Model
Oryx-7B	Huggingface	7B	Oryx-ViT	Qwen-2-7B	Oryx-7B-Image
Oryx-34B	Huggingface	34B	Oryx-ViT	Yi-1.5-34B	Oryx-34B-Image
Oryx-1.5-7B	Huggingface	7B	Oryx-ViT	Qwen-2.5-7B	Oryx-1.5-7B-Image
Oryx-1.5-32B	Huggingface	32B	Oryx-ViT	Qwen-2.5-32B	Oryx-1.5-32B-Image

Generation Demo

You can try the generation results of our strong Oryx model with the following steps:

1. Download the Oryx model from our huggingface collections.

2. Download the Oryx-ViT vision encoder.

3. Replace the path for "mm_vision_tower" in the config.json with your local path for Oryx-ViT.

4. Modify the model path and run the inference script with your own video to test our model.

python inference.py

Evaluation

You can evaluate our model with the following steps:

1. Download the Oryx model from our huggingface collections.

2. Download the Oryx-ViT vision encoder.

3. Replace the path for "mm_vision_tower" in the config.json with your local path for Oryx-ViT.

4. Install the provided lmms-eval folder.

cd ./lmms-eval
pip install -e .

4. Modify the model path and run the evaluation script to test our model.

bash ./scripts/eval_image.sh
bash ./scripts/eval_video.sh

Training Instructions

Installation

1. Clone this repository:

git clone https://github.com/Oryx-mllm/oryx
cd oryx

2. Install the required package:

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

Preparation

3. Prepare training data:

Please download training data from our huggingface.

Modify the DATA and FOLDER arguments in the training scripts to your save folder.

DATA="PATH/TO/Oryx-SFT-Data/data.json"
FOLDER="PATH/TO/Oryx-SFT-Data"

If you are interested in our long-form training data, you can download movienet_data.json and movienet_patch and mix appropriate quantity (we recommand 30k) with the main training data.

Training

4. Training your own model:

Modify the following lines in the scripts at your own environments:

export PYTHONPATH=/PATH/TO/oryx:$PYTHONPATH
VISION_TOWER='oryx_vit:PATH/TO/oryx_vit_new.pth'
DATA="PATH/TO/Oryx-SFT-DATA/data.json"
MODEL_NAME_OR_PATH="PATH/TO/7B_MODEL"

Scripts for training Oryx-7B

bash scripts/train_oryx_7b.sh

Scripts for training Oryx-34B

bash scripts/train_oryx_34b.sh

Citation

If you find it useful for your research and applications, please cite our paper using this BibTeX:

@article{liu2024oryx,
title={Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution},
author={Liu, Zuyan and Dong, Yuhao and Liu, Ziwei and Hu, Winston and Lu, Jiwen and Rao, Yongming},
journal={arXiv preprint arXiv:2409.12961},
year={2024}
}

Acknowledgement

Our codebase is conducted on LLaVA
Thanks to lmms-eval team, for building such a usefule evaluation system!

Contributors

25 commits

13 commits

4 commits

3 commits

Similar repos

liuzuyan/oryx

MLLM for On-Demand Spatial-Temporal Understanding at Arbitrary Resolution

THUdyh/Oryx-1.5-7B

Model

Oryx-1.5-7B

THUdyh/Oryx-ViT

Model

Oryx-ViT

THUdyh/Oryx-34B

Model

Oryx-34B

THUdyh/Oryx-1.5-32B

Model

Oryx-1.5-32B

ATH-MaaS/Ovis

A novel Multimodal Large Language Model (MLLM) architecture, designed to structurally align visual…

1,522

AIDC-AI/Ovis

A novel Multimodal Large Language Model (MLLM) architecture, designed to structurally align visual…

1,522

THUdyh/Oryx-7B

Model

Oryx-7B

Languages

Python

99.6%

Oryx-mllm/Oryx

[ICLR 2025] MLLM for On-Demand Spatial-Temporal Understanding at Arbitrary Resolution

Python

329

47 commits

updated Jul 4, 2025

oryx-mllm.github.io

See the code

README

Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution

Zuyan Liu^*,1,2 Yuhao Dong^*,2,3 Ziwei Liu³ Winston Hu² Jiwen Lu^1,✉ Yongming Rao^2,1,✉

¹Tsinghua University ²Tencent ³S-Lab, NTU

^* Equal Contribution^✉ Corresponding Author

Project Page:

arXiv Paper:

Demo by Gradio:

Model Checkpoints:

Oryx SFT Data:

📢 News

[23/1/2025] 🎉Oryx has been accepted to ICLR 2025!
[23/10/2024] 🎉Oryx-1.5 Series is released! Oryx-1.5 includes 7B and 32B variants. We achieve stronger performance on all the benchmarks! Check our results at VideoMME Leaderboard and the updated arXiv paper.
[30/09/2024] 📊Oryx video data for SFT is released!
[26/09/2024] 🎨Try out our online demo with Oryx-7B for image/video understanding!
[24/09/2024] 🚀 Oryx-34B is now available at VideoMME Leaderboard, Oryx-34B achieves best accuracy among <40B MLLMs.
[23/09/2024] 🔥 Oryx ranks 1st on MLVU, surpassing GPT-4o. Stay tuned for more results!
[20/09/2024] 🔥 🚀Introducing Oryx! The Oryx models (7B/34B) support on-demand visual perception, achieve new state-of-the-art performance across image, video and 3D benchmarks, even surpassing advanced commercial models on some benchmarks.
- [Paper]: Detailed introduction of on-demand visual perception, including native resolution perception and dynamic compressor!
- [Checkpoints]: Try our advanced model on your own.
- [Scripts]: Start training models with customized data.

🐐 Introducing Oryx

Main idea of On-Demand Multimodal Understanding

Overview of Oryx Architecture

✅ TODO List

Release all the model weights.
Release OryxViT model.
Demo code for generation.
All the training and inference code.
Evaluation code for image, video and 3D multi-modal benchmark.
Oryx SFT Data.
Oryx Gradio chatbox.
Enhanced Oryx model with latest LLM base models and better SFT data.
Introducing our explorations for OryxViT.

📃 Main Results

Results on General Temporal Understanding

Results on Long-Form Temporal Understanding

Results on Image Understanding

Results on 3D Understanding

Model Zoo

We provide our checkpoints at Huggingface

Model	Link	Size	Visual Encoder	LLM-Type	Intermediate Model
Oryx-7B	Huggingface	7B	Oryx-ViT	Qwen-2-7B	Oryx-7B-Image
Oryx-34B	Huggingface	34B	Oryx-ViT	Yi-1.5-34B	Oryx-34B-Image
Oryx-1.5-7B	Huggingface	7B	Oryx-ViT	Qwen-2.5-7B	Oryx-1.5-7B-Image
Oryx-1.5-32B	Huggingface	32B	Oryx-ViT	Qwen-2.5-32B	Oryx-1.5-32B-Image

Generation Demo

You can try the generation results of our strong Oryx model with the following steps:

1. Download the Oryx model from our huggingface collections.

2. Download the Oryx-ViT vision encoder.

3. Replace the path for "mm_vision_tower" in the config.json with your local path for Oryx-ViT.

4. Modify the model path and run the inference script with your own video to test our model.

python inference.py

Evaluation

You can evaluate our model with the following steps:

1. Download the Oryx model from our huggingface collections.

2. Download the Oryx-ViT vision encoder.

3. Replace the path for "mm_vision_tower" in the config.json with your local path for Oryx-ViT.

4. Install the provided lmms-eval folder.

cd ./lmms-eval
pip install -e .

4. Modify the model path and run the evaluation script to test our model.

bash ./scripts/eval_image.sh
bash ./scripts/eval_video.sh

Training Instructions

Installation

1. Clone this repository:

git clone https://github.com/Oryx-mllm/oryx
cd oryx

2. Install the required package:

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

Preparation

3. Prepare training data:

Please download training data from our huggingface.

Modify the DATA and FOLDER arguments in the training scripts to your save folder.

DATA="PATH/TO/Oryx-SFT-Data/data.json"
FOLDER="PATH/TO/Oryx-SFT-Data"

If you are interested in our long-form training data, you can download movienet_data.json and movienet_patch and mix appropriate quantity (we recommand 30k) with the main training data.

Training

4. Training your own model:

Modify the following lines in the scripts at your own environments:

export PYTHONPATH=/PATH/TO/oryx:$PYTHONPATH
VISION_TOWER='oryx_vit:PATH/TO/oryx_vit_new.pth'
DATA="PATH/TO/Oryx-SFT-DATA/data.json"
MODEL_NAME_OR_PATH="PATH/TO/7B_MODEL"

Scripts for training Oryx-7B

bash scripts/train_oryx_7b.sh

Scripts for training Oryx-34B

bash scripts/train_oryx_34b.sh

Citation

If you find it useful for your research and applications, please cite our paper using this BibTeX:

@article{liu2024oryx,
title={Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution},
author={Liu, Zuyan and Dong, Yuhao and Liu, Ziwei and Hu, Winston and Lu, Jiwen and Rao, Yongming},
journal={arXiv preprint arXiv:2409.12961},
year={2024}
}