How Open‑Sora 2.0 Achieves SOTA Video Generation with Only $200K Training Cost

Open‑Sora 2.0 Overview

Open‑Sora 2.0 is an open‑source video generation model with 11 B parameters. It was trained on 224 GPUs with a budget of ≈ $200 k, achieving commercial‑grade quality comparable to proprietary models such as HunyuanVideo and Step‑Video (30 B).

Performance Benchmarks

VBench and human‑preference evaluations show Open‑Sora 2.0 matches or exceeds closed‑source models. The VBench gap to OpenAI Sora decreased from 4.52 % to 0.69 %, and it surpasses Tencent’s HunyuanVideo on the same benchmark.

Model Architecture

The model builds on Open‑Sora 1.2, retaining a 3D auto‑encoder and Flow‑Matching training framework, with the following enhancements:

3D full‑attention mechanism for higher visual fidelity.

MMDiT backbone for improved text‑to‑video alignment.

Scale increased from 1 B to 11 B parameters.

Initialization from the open‑source FLUX image‑to‑video model to reduce training cost.

Cost‑Effective Training Strategies

Four optimizations reduce training expense:

Strict multi‑stage data filtering to ensure high‑quality training data.

Primary training at low resolution (256 px) which reduces token count from ~8 k (768 px) to ~80 k and avoids quadratic attention cost.

Image‑to‑video pre‑training, which converges faster than direct high‑resolution video training.

Efficient parallel training stack based on ColossalAI, including sequence parallelism, ZeRO‑DP, gradient checkpointing, automatic recovery, optimized data loading, asynchronous checkpoint saving, and operator‑level optimizations.

High‑Compression Auto‑Encoder

A 4 × 32 × 32 compression auto‑encoder reduces single‑GPU generation time for a 768 px × 5 s video from ~30 minutes to under 3 minutes (≈10× speed‑up). Training techniques include:

Residual connections in the video up‑down‑sampling module for stable reconstruction.

Distillation‑based optimization to improve latent representation.

Initialization from a pre‑trained high‑quality model to lower data and time requirements.

Open‑Source Resources

Model weights, training code, and distributed training pipeline are available at:

https://github.com/hpcaitech/Open-Sora

Technical report: https://github.com/hpcaitech/Open-Sora-Demo/blob/main/paper/Open_Sora_2_tech_report.pdf