Search papers, labs, and topics across Lattice.
The paper identifies that inaccurate self-conditioning in few-step diffusion language models leads to a significant approximation gap, hindering sample quality. To mitigate this, they introduce a training framework, FastDiSS, that perturbs the self-conditioning signal during training to match inference noise, enhancing robustness to prior estimation errors. They further incorporate a token-level noise-awareness mechanism to prevent training saturation, achieving up to 400x faster inference speeds compared to standard continuous diffusion models while maintaining competitive performance.
Diffusion language models can achieve 400x faster inference without sacrificing quality by correcting self-conditioning errors that compound in few-step sampling.
Self-conditioning has been central to the success of continuous diffusion language models, as it allows models to correct previous errors. Yet its ability degrades precisely in the regime where diffusion is most attractive for deployment: few-step sampling for fast inference. In this study, we show that when models only have a few denoising steps, inaccurate self-conditioning induces a substantial approximation gap; this mistake compounds across denoising steps and ultimately dominate the sample quality. To address this, we propose a novel training framework that handles these errors during learning by perturbing the self-conditioning signal to match inference noise, improving robustness to prior estimation errors. In addition, we introduce a token-level noise-awareness mechanism that prevents training from saturation, hence improving optimization. Extensive experiments across conditional generation benchmarks demonstrate that our framework surpasses standard continuous diffusion models while providing up to 400x faster inference speed, and remains competitive against other one-step diffusion frameworks.
