Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation

A research team led by Xingtong Ge has introduced 'Salt,' a novel training method that dramatically accelerates AI video generation while preserving quality. The technique combines Self-Consistent Distribution Matching (SC-DMD) with Cache-Aware Training to enable models to produce high-quality videos in just 2-4 Neural Function Evaluations (NFEs), making real-time generation feasible. Traditional distillation methods either produce over-smoothed results or suffer from 'drift' where video quality degrades over time, but Salt's approach explicitly regularizes how denoising updates compose across timesteps.

The method's key innovation is treating the KV cache - a memory mechanism in transformer models - as a quality-parameterized condition during training. This cache-distribution-aware approach applies SC-DMD over multi-step rollouts and introduces a feature alignment objective that steers low-quality outputs toward high-quality references. The researchers demonstrated Salt's effectiveness across multiple architectures, including non-autoregressive backbones like Wan 2.1 and autoregressive real-time paradigms like Self Forcing, showing consistent improvements in low-NFE video generation quality while remaining compatible with diverse memory mechanisms.

Salt addresses a critical bottleneck in deploying AI video generation for interactive applications, where inference speed directly impacts user experience. By achieving high-quality results with dramatically fewer computational steps, the method opens possibilities for real-time video editing, gaming, and interactive media creation that were previously constrained by generation latency. The team has made their source code publicly available, potentially accelerating adoption across the AI video generation ecosystem.