Mira: A Mini-step Towards Sora-like Long Video Generation

Zhaoyang Zhang¹, Ziyang Yuan¹, Xuan Ju¹, Yiming Gao¹, Xintao Wang^1#, Chun Yuan, Ying Shan¹,
¹ARC Lab, Tencent PCG ^*Equal contribution ^#Project lead

We introduce Mira (Mini-Sora), an initial foray into the realm of high-quality, long-duration video generation in the style of Sora. Mira stands out from existing text-to-video (T2V) generation frameworks in several key ways:

• Extended sequence length: While most frameworks are limited to generating short videos (2 seconds / 16 frames), Mira is designed to produce significantly longer sequences, potentially lasting 10 seconds, 20 seconds, or more.

• Enhanced dynamics: Mira has the capability to create videos with rich dynamics and intricate motions, setting it apart from the more static outputs of current video generation technologies.

• Strong 3D consistency: Despite the intricate dynamics and object interactions, Mira ensures the 3D integrity of objects is preserved throughout the video, avoiding noticeable distortions.

Please acknowledge that our work on Mira is in the experimental phase. There are several areas where Sora still significantly outperforms Mira and other open-source T2V frameworks, including:

• Interactive objects and environments: Sora supports the generation of videos where objects and surroundings engage in dynamic interactions, adding a layer of complexity and realism.

• Sustained object consistency: Sora maintains consistent object shapes, even when they temporarily exit and re-enter the frame, ensuring continuity and coherence.

The Mira project is our endeavor to investigate and refine the entire data-model-training pipeline for Sora-like, lightweight T2V frameworks, and to preliminarily demonstrate the aforementioned Sora characteristics. Our goal is to foster innovation and democratize the field of content creation, paving the way for more accessible and advanced video generation tools.

Results

5s 768x480

https://github.com/mira-space/Mira/assets/163223899/5e7d74d3-82a4-4a94-bfc1-9140b7929c50

10s 384×240

https://github.com/mira-space/Mira/assets/13939478/4de6aade-4eca-4291-bcc6-950c7b44c981

Each individual video can be downloaded from here.

20s 128×80

https://github.com/mira-space/Mira/assets/13939478/9f274503-9715-4d2a-a262-10113c4df78f

📰 Updates

Stay tuned! We are actively working on this project. Expect a steady stream of updates as we expand our dataset, enhance our annotation processes, and refine our model checkpoints. Keep an eye out for these upcoming updates, as we continue to make strides in our project's development.

[2024.07.11] 🔥 We're glad to announce the release of Mira-v1 and MiraData-v1! The full version of the MiraData-v1 datasets is now available, along with the corresponding technical report](https://arxiv.org/abs/2407.06358v1). Additionally, we have updated the MiraDiT model to improve quality, now supporting resolutions up to 768x480 and durations up to 10 seconds using the updated data. This version also includes an optional post-processing feature for video interpolation and enhancement, leveraging the RIFE](https://github.com/hzwer/ECCV2022-RIFE) framework.

[2024.04.01] 🔥 We're delighted to announce the release of Mira and MiraData-v0. This release offers a comprehensive open-source suite for data annotation and training pipelines, specifically tailored for the creation of long-duration videos with dynamic content and consistent quality. Our provided codes and checkpoints empower users to generate videos up to 20 seconds in 128x80 resolution and 10 seconds in 384x240 resolution. Dive into the future of video generation with Mira!

Installation

## create a conda enviroment
conda update -n base -c defaults conda 
conda create -y -n mira python=3.8.5 
source activate mira 

## install dependencies
pip install torch==2.0 torchvision torchaudio decord==0.6.0  \
einops==0.3.0  imageio==2.9.0 \
numpy omegaconf==2.1.1 opencv_python pandas \
Pillow==9.5.0 pytorch_lightning==1.9.0 PyYAML==6.0 setuptools==65.6.3  \
torchvision tqdm==4.65.0 transformers==4.25.1 moviepy av  tensorboardx \
&& pip install  timm scikit-learn  open_clip_torch==2.22.0 kornia simplejson easydict pynvml rotary_embedding_torch==0.3.1 triton  cached_property  \
&& pip install xformers==0.0.18 \
&& pip install taming-transformers fairscale deepspeed  diffusers

Training

Checkpoints

Please download the above checkponits in our huggingface page (Mira-V0) and huggingface page (Mira-V0).

• Please note that some Mira-v1 models are larger in size because they were trained using FP32 precision to ensure training stability.

Finetuning the Mira-v0 model on 768x480 resolution.

• Add path to your datasets and the pretrain models in config_768v1_5s_mira.yaml.
• Then conduct the following commands:

## activate envrionment
conda activate mira

## Run training
bash configs/Mira/run_768v1_mira.sh 0

Finetuning the Mira-v0 model on 384x240 resolution.

• Add path to your datasets and the pretrain models in config_384v1_10s_mira.yaml.
• Then conduct the following commands:

## activate envrionment
conda activate mira

## Run training
bash configs/Mira/run_384v1_mira.sh 0

Inference

Evaluate the Mira-v1 model on 768x480 resolution.

• Add path to your model checkponits in run_text2video_768.sh.
• Add your test prompts in test_prompt.txt.
• Then conduct the following commands:

## activate envrionment
conda activate mira

## Run inference
bash configs/inference/run_text2video_768.sh

Evaluate the Mira-v1 model on 384x240 resolution.

• Add path to your model checkponits in run_text2video_384.sh.
• Add your test prompts in test_prompt.txt.
• Then conduct the following commands:

## activate envrionment
conda activate mira

## Run inference
bash configs/inference/run_text2video_384.sh

Licence

Mira is under the GPL-v3 Licence and is supported for commercial usage. If you need a commercial license for YOLO-World, please feel free to contact us.