Researchers have developed SWIFT, a novel self-guided training method for diffusion models that enhances stronger models using a frozen, weaker denoiser. This approach extrapolates the clean regression target away from the weak denoiser's prediction, amplifying underfit components without increasing inference costs. SWIFT can utilize an early training checkpoint of the weak guide, which is discarded after training, and supports both conditional and unconditional generation. Experiments on CIFAR-10 and ImageNet-256 datasets show improvements in model performance and efficiency, with one configuration achieving a better FID score using half the training compute compared to standard methods. AI
IMPACT This method could lead to more efficient training of high-quality diffusion models, reducing computational costs and accelerating development.
RANK_REASON The cluster contains a research paper detailing a new method for training diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
