Open-Sora-Plan

Open-Sora-Plan is a community effort led by Peking University's Yuan Lab and Tuozhan AIGC Lab to reproduce Sora-class video generation in the open, and it is distinct from HPC-AI Tech's similarly named Open-Sora project. The project develops text-to-video and image-to-video models on a diffusion transformer stack with its own components: WFVAE, a wavelet-enhanced video autoencoder with 8x8x8 downsampling, and SUV, a sparse 3D attention diffusion transformer the team reports as roughly 35 percent faster than dense attention at similar quality. Version 1.5.0 is an 8 billion parameter model trained on 40 million video samples, generating up to 121 frames at 576x1024 resolution with variable resolution and duration through bucket training; this release runs on Huawei Ascend 910 accelerators, with a GPU build planned. Contributors include Huawei and Pengcheng Laboratory, and code and models are released under an Apache license.