Open-Sora 1.2 by HPC AI Tech: Transforming Video Generation With Advanced, Open-Source Video Generation and Compression

Open-Sora, an initiative by HPC AI Tech, is a great innovation in democratizing efficient video production. By embracing open-source principles, Open-Sora aims to make advanced video generation techniques accessible to everyone, fostering innovation, creativity, and inclusivity in content creation.

Open-Sora 1.0 and 1.1

Open-Sora 1.0 laid the groundwork for this project, offering a full pipeline for video data preprocessing, training, and inference. It supports generating videos up to 2 seconds long at 512×512 resolution with a minimal training cost. Following this, Open-Sora 1.1 expanded capabilities to support 2-15 second videos, ranging from 144p to 720p, and various aspect ratios. It introduced a comprehensive video processing pipeline, including scene cutting, filtering, and captioning, making it easier for users to build their video datasets.

Key Features of Open-Sora

Open-Sora aims to simplify the complexities of video generation by providing a streamlined and user-friendly platform. Its primary features include:

  • Text-to-Video Generation: Users can generate videos based on textual descriptions.
  • Image-to-Video Generation: This feature allows images to be transformed into video sequences.
  • Video-to-Video Translation: Users can convert one video format to another with ease.

Open-Sora 1.2 Enhancements

Open-Sora 1.2 introduces several notable improvements over its predecessors. It includes a 3D-VAE model, rectified flow, and score conditioning, significantly enhancing video quality. The update also focuses on better data handling and multi-stage training, ensuring the model can handle more complex tasks efficiently.

  1. Video Compression Network: The new version incorporates OpenAI’s Sora, which improves video compression by reducing temporal dimensions without sacrificing frame rates. This results in smoother, high-quality video output.
  2. Rectified Flow Training: Adopting techniques from the latest diffusion models, Open-Sora 1.2 includes rectified flow training, enhancing the performance and quality of generated videos.
  3. Evaluation Metrics: Open-Sora 1.2 supports advanced evaluation metrics like validation loss, VBench score, and VBench-i2v score, ensuring comprehensive assessment during the training process. The improvements in evaluation can be seen in the higher quality and semantic scores compared to previous versions.

The training process for Open-Sora 1.2 remains similar to earlier versions but with enhanced configurations. The model is trained on over 30 million data points, utilizing 80,000 GPU hours supporting various video resolutions and aspect ratios. The command line for inference supports multiple configurations, including text-to-video and image-to-video generation.

Open-Sora 1.2 provides model weights and a detailed installation guide, ensuring users can deploy the system easily. The installation process supports various CUDA versions and includes dependencies for data preprocessing, VAE, and model evaluation.


Open-Sora 1.2 by HPC AI Tech is a robust and innovative solution for video generation, incorporating state-of-the-art techniques and open-source accessibility. With its continuous improvements and community-driven approach, Open-Sora is poised to revolutionize content creation.


 | Website

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.

🐝 Join the Fastest Growing AI Research Newsletter Read by Researchers from Google + NVIDIA + Meta + Stanford + MIT + Microsoft and many others...