Open-Sora 1.1 Released: Major Improvements in Video Quality and Duration 203

Today, we are excited to introduce Open-Sora-Plan v1.1.0, which significantly improves video generation quality and duration, as well as text control capabilities. Compared with the previous Open-Sora-Plan v1.1.0 version, the improvements include:

Better Compressed Visual Representations: We have optimized the CausalVideoVAE architecture, which now delivers stronger performance and higher inference efficiency.

Higher Generation Quality and Longer Duration: We have leveraged higher-quality visual data and captions, along with ShareGPT4Video's ability to annotate long videos, enabling us to generate higher-quality and longer videos.

In addition to the performance improvements, Open-Sora-Plan v1.1.0 retains the minimalist design and data efficiency of v1.0.0. Notably, we found that v1.1.0 exhibits performance similar to that of the Sora base model, indicating that the evolution of our version is consistent with the scaling laws demonstrated by Sora.

We have open-sourced Open-Sora-Plan to promote the future development of community video generation. The code, data, and models will be made publicly available.

Demo: Hugging Face demo here.

Code: All training scripts and example scripts.

Model: Diffusion Model and CausalVideoVAE here.

Data: Raw videos and subtitles here.

GitHub: https://github.com/PKU-YuanGroup/Open-Sora-Plan/tree/main

Online Experience: https://huggingface.co/spaces/LanguageBind/Open-Sora-Plan-v1.1.0

Open-Sora-Plan v1.1.0 is a transformer-based text-to-video diffusion system, trained on T5-based text embeddings.