🚀 Introducing scale and spatial Markovian assumptions into VAR — achieving faster generation with reduced memory consumption.

[NuanBaobao]

Thanks to the great work on VAR! 🙌
Building upon it, we introduce MVAR, which incorporates scale and spatial Markovian assumptions into visual autoregressive modeling.
MVAR achieves a 1.7× speedup and 3× reduction in GPU memory usage, enabling efficient training on eight RTX 4090 GPUs.

📄 Paper: https://arxiv.org/abs/2505.1274
🔗 GitHub: https://github.com/LabShuHangGU/MVAR