Swift begins with a weak denoiser, which can be an early checkpoint from ordinary diffusion training.
Biased training targets help diffusion models learn stronger denoisers
Swift trains a diffusion model to extrapolate away from a weaker denoiser, improving generation while retaining a single standard sampler at inference.
Chinese Tech
Yanjin He · Cuifeng Shen · Wanzhou Lei · Maohua Li · Tao Lan · Kan Liu · +2 more
University of Michigan, Ann Arbor · Alibaba Group · University of California, Berkeley · Nanjing University
Research Digest··3 min read
He and colleagues introduce Swift, a training method that uses a frozen, undertrained denoiser to modify the regression targets seen by a stronger model.
§