Found first: a primary source the press has not covered yet.
A paper from eight researchers introduces SWIFT, a training method for diffusion models that surpasses a strong baseline's best FID using roughly half the total training compute, with no added cost at inference. On ImageNet-256, it reaches FID 1.15 without guidance and FID 0.99 with classifier-free guidance on the RAEv2 representation. The paper is on arXiv.
What the source says
The authors are Yanjin He, Cuifeng Shen, Wanzhou Lei, Maohua Li, Tao Lan, Kan Liu, Hanlin Tang, and Qing Qu; the paper does not specify their institutional affiliations on the abstract page. SWIFT works by freezing an early-checkpoint diffusion model, a "weak denoiser," and using its predictions to extrapolate clean regression targets for the main model during training. The weak guide receives no gradients and is discarded once training is complete, so inference pipelines are unchanged. The compute comparison includes the cost of preparing the guide. Theoretical support comes from a spectral analysis of finite-time population gradient flow in a reproducing kernel Hilbert space, with a proof that a range of extrapolation strengths yields strictly lower posterior-mean error than standard training at the same step count.
Why it matters
Training competitive diffusion models is one of the more expensive operations in modern generative modeling. A method that reaches the same quality ceiling at roughly half the compute, with no inference overhead added, is directly applicable to teams building image and video generators at scale. The savings are entirely on the training side, which means existing deployment infrastructure needs no changes. The method works across model scales and supports both conditional and unconditional generation without label dropout, so it is not limited to a narrow experimental setup.