Reproduce “IntMeanFlow: Few-step Speech Generation with Integral Velocity Distillation” on OpenResearch

About this paper

Wei Wang, Rong Cao, Yi Guo, Zhengyang Chen, Kuan Chen, Yuanyuan Huo

ByteDance researchers developed IntMeanFlow, a distillation-based framework for few-step speech generation that enables high-quality synthesis while significantly reducing inference steps. The approach overcomes computational and stability issues of previous MeanFlow models by learning integral velocity without Jacobian-vector products and integrates an optimal sampling strategy, achieving up to 10x acceleration in Real-Time Factor.

Key insights

• Learning integral velocity through distillation effectively bypasses the computational and stability challenges associated with Jacobian-vector products in MeanFlow, enabling efficient few-step generation for TTS.
• Optimizing the distribution of a fixed number of inference steps, rather than just the number of steps, can substantially improve speech synthesis quality without increasing the computational burden.
• A capable teacher model, even if larger, can effectively instill high performance in a more compact student model during the distillation process, enabling efficient yet high-fidelity speech generation.

Summary by alphaXiv.