2026-10-21 · America/Los_Angeles · 论文 · #20

SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation

Training high-quality diffusion models is computationally expensive, and inference-time guidance often adds further sampling cost. In this work, we introduce Swift, a simple yet effective self-guided training method that uses a frozen weak denoiser to improve a stronger model without adding inference overhead. For each noisy input, Swift extrapolates the clean regression target away from the weak denoiser's prediction, amplifying components that remain underfit. The weak guide can be an early training checkpoint, requires no backpropagation during main-model training, and is discarded…

热度 44.0 / 100；排名与评分保留该期记录。