🔆Introduction

⚡️ Hummingbird is a series of advanced video generation models developed by the AMD AIG team and trained on AMD Instinct™ MI250 GPUs. It includes text-to-video models, image-to-video models, and image/video super-resolution models. With only 0.9B parameters, the Hummingbird model demonstrates exceptional efficiency. For text-to-video tasks, it can generate text-aligned videos in just 1.87 seconds using 4 steps on an MI250 GPU. For image-to-video tasks, it takes only 11 seconds to produce high-quality 4K videos.

📝 Change Log

• [2026.01.09]: 🔥🔥Release the full code and pre-trained weight of HummingbirdXT!
• [2026.01.08]: 🔥🔥Our new model Hummingbird-XT has been released, this is the technical report link: : Bridging the Last Mile: Deploying Hummingbird-XT for Efficient Video Generation on AMD Consumer-Grade Platforms!
• [2025.08.03]: 🔥🔥Release Hummingbird Image-to-Video Technical Report!
• [2025.07.30]: 🔥🔥Release pretrained Image-to-Video model and VSR model, and their training and inference code!
• [2025.03.24]: 🔥🔥Release AMD-Hummingbird: Towards an Efficient Text-to-Video Model Paper!
• [2025.02.28]: 🔥🔥Release Hummingbird Text-to-Video Technical Report!
• [2025.02.26]: 🔥🔥Release pretrained Text-to-Video models, training and inference code!

🚀Getting Started

Installation

Conda

conda create -n AMD_Hummingbird python=3.10
conda activate AMD_Hummingbird
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/rocm6.1
pip install -r requirements.txt

For rocm flash-attn, you can install it by this link.

git clone https://github.com/ROCm/flash-attention.git
cd flash-attention
python setup.py install

It will take about 1.5 hours to install.

Docker

First, you should use docker pull to download the image.

docker pull rocm/vllm:rocm6.2_mi300_ubuntu20.04_py3.9_vllm_0.6.4

Second, you can use docker run to run the image, for example:

docker run \
        -v "$(pwd):/workspace" \
        --device=/dev/kfd \
        --device=/dev/dri \
        -it \
        --network=host \
        --name hummingbird \
        rocm/vllm:rocm6.2_mi300_ubuntu20.04_py3.9_vllm_0.6.4

When you in the container, you can use pip to install other dependencies:

pip install -r requirements.txt

Example Usage

Text-to-Video

Download the Unet pretrained checkpoint from Hummingbird-Text-to-Video. Run below command to generate videos:

# for 0.7B model
python inference_command_config_07B.py

# for 0.9B model
python inference_command_config_09B.py

Image-to-Video

Download the Image-to-Video pretrained checkpoint from Hummingbird-Image-to-Video. Run below command to generate videos:

cd i2v
sh run_hummingbird.sh

Image/Video Super-Resolution

Download SR pretrained checkpoint from Hummingbird-Image-to-Video. Run below command to generate high-resolution videos:

cd VSR
sh inference_videos.sh

💥Pre-training

Data Preparation

# VQA
cd data_pre_process/DOVER
sh run.sh

Then you can get a score table for all video qualities, sort according to the table, and remove low-scoring videos.

# Remove Dolly Zoom Videos
cd data_pre_process/VBench
sh run.sh 

According to the motion smoothness score csv file, you can remove low-scoring videos.

Training

Text-to-video

cd acceleration/t2v-turbo

# for 0.7 B model
sh train_07B.sh

# for 0.9 B model
sh train_09B.sh

Image/Video Super-Resolution

Firstly, you should train the Realesrnet model:

cd VSR
# for realesrnet model
sh train_realesrnet.sh

And you will get the trained checkpoint of Realesrnet, then you can train the Realesrgan model:

cd VSR
# for realesrgan model
sh train_realesrgan.sh

🤗Resources

Pre-trained models

• Text-to-Video: Hummingbird-Text-to-Video
• Image-to-Video: Hummingbird-Image-to-Video
• Image/Video Super-Resolution: Hummingbird-SR

AMD Blogs

Please refer to the following blogs to get started with using these techniques on AMD GPUs:

• PyTorch Fully Sharded Data Parallel (FSDP) on AMD GPUs with ROCm™
• Accelerating Large Language Models with Flash Attention on AMD GPUs
• Accelerate PyTorch Models using torch.compile on AMD GPUs with ROCm™
• Introducing the First AMD 1B Language Models: AMD OLMo

❤️Acknowledgement

Our codebase builds on VideoCrafter2, DynamicCrafter, T2v-Turbo, Real-ESRGAN .Thanks the authors for sharing their awesome codebases!

📋Citations

Feel free to cite our Hummingbird models and give us a star⭐, if you find our work helpful :)

@article{isobe2025amd,
  title={AMD-Hummingbird: Towards an Efficient Text-to-Video Model},
  author={Isobe, Takashi and Cui, He and Zhou, Dong and Ge, Mengmeng and Li, Dong and Barsoum, Emad},
  journal={arXiv preprint arXiv:2503.18559},
  year={2025}
}