StreamingT2V

This repository is the official implementation of StreamingT2V.

StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text
Roberto Henschel*, Levon Khachatryan*, Daniil Hayrapetyan*, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, Humphrey Shi

StreamingT2V is an advanced autoregressive technique that enables the creation of long videos featuring rich motion dynamics without any stagnation. It ensures temporal consistency throughout the video, aligns closely with the descriptive text, and maintains high frame-level image quality. Our demonstrations include successful examples of videos up to 1200 frames, spanning 2 minutes, and can be extended for even longer durations. Importantly, the effectiveness of StreamingT2V is not limited by the specific Text2Video model used, indicating that improvements in base models could yield even higher-quality videos.

News

[03/21/2024] Paper StreamingT2V released!
[04/05/2024] Code and model released!
[04/06/2024] The first version of our huggingface demo released!

Setup

Clone this repository and enter:

git clone https://github.com/Picsart-AI-Research/StreamingT2V.git
cd StreamingT2V/

Install requirements using Python 3.10 and CUDA >= 11.6

conda create -n st2v python=3.10
conda activate st2v
pip install -r requirements.txt

(Optional) Install FFmpeg if it's missing on your system

conda install conda-forge::ffmpeg

Download the weights from HF and put them into the t2v_enhanced/checkpoints directory.

Inference

For Text-to-Video

cd t2v_enhanced
python inference.py --prompt="A cat running on the street"

To use other base models add the --base_model=AnimateDiff argument. Use python inference.py --help for more options.

For Image-to-Video

cd t2v_enhanced
python inference.py --image=../__assets__/demo/fish.jpg --base_model=SVD

Inference Time

ModelscopeT2V as a Base Model

Number of Frames	Inference Time for Faster Preview (256x256)	Inference Time for Final Result (720x720)
24 frames	40 seconds	165 seconds
56 frames	75 seconds	360 seconds
80 frames	110 seconds	525 seconds
240 frames	340 seconds	1610 seconds (~27 min)
600 frames	860 seconds	5128 seconds (~85 min)
1200 frames	1710 seconds (~28 min)	10225 seconds (~170 min)

AnimateDiff as a Base Model

Number of Frames	Inference Time for Faster Preview (256x256)	Inference Time for Final Result (720x720)
24 frames	50 seconds	180 seconds
56 frames	85 seconds	370 seconds
80 frames	120 seconds	535 seconds
240 frames	350 seconds	1620 seconds (~27 min)
600 frames	870 seconds	5138 seconds (~85 min)
1200 frames	1720 seconds (~28 min)	10235 seconds (~170 min)

SVD as a Base Model

Number of Frames	Inference Time for Faster Preview (256x256)	Inference Time for Final Result (720x720)
24 frames	80 seconds	210 seconds
56 frames	115 seconds	400 seconds
80 frames	150 seconds	565 seconds
240 frames	380 seconds	1650 seconds (~27 min)
600 frames	900 seconds	5168 seconds (~86 min)
1200 frames	1750 seconds (~29 min)	10265 seconds (~171 min)

All measurements were conducted using the NVIDIA A100 (80 GB) GPU. Randomized blending is employed when the frame count surpasses 80. For Randomized blending, the values for chunk_size and overlap_size are set to 112 and 32, respectively.

Gradio

The same functionality is also available as a gradio demo

cd t2v_enhanced
python gradio_demo.py

Results

Detailed results can be found in the Project page.

License

Our code is published under the CreativeML Open RAIL-M license.

We include ModelscopeT2V, AnimateDiff, SVD in the demo for research purposes and to demonstrate the flexibility of the StreamingT2V framework to include different T2V/I2V models. For commercial usage of such components, please refer to their original license.

BibTeX

If you use our work in your research, please cite our publication:

@article{henschel2024streamingt2v,
  title={StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text},
  author={Henschel, Roberto and Khachatryan, Levon and Hayrapetyan, Daniil and Poghosyan, Hayk and Tadevosyan, Vahram and Wang, Zhangyang and Navasardyan, Shant and Shi, Humphrey},
  journal={arXiv preprint arXiv:2403.14773},
  year={2024}
}

tyfeld/StreamingT2V