SWIFT halves ImageNet-256 diffusion training via weak denoiser

A new arXiv preprint proposes a diffusion training recipe called Swift that claims to reach the same image-quality ceiling using roughly half the compute. The method keeps a frozen weak denoiser alongside the main model and, for each noisy input, pushes the clean regression target away from the weak model's prediction, "amplifying components that remain underfit," the paper reports. The weak guide is discarded after training, so inference overhead is unchanged.

The paper, by Yanjin He, Cuifeng Shen, Wanzhou Lei and five collaborators, states that in "one ImageNet-256 configuration, Swift surpasses the baseline's best FID using roughly half the total training compute, including guide preparation and training overhead." On the RAEv2 representation the authors report FID 1.15 without inference-time guidance and 0.99 with guidance.

The claim rests on a spectral analysis of finite-time population gradient flow in a reproducing kernel Hilbert space, with experiments on CIFAR-10 and ImageNet-256 across model scales. The abstract does not name the specific baseline model the half-compute figure is measured against, publish absolute GPU-hour counts, or describe how the early-stopped weak-guide checkpoint is chosen, so the headline number is a single-paper assertion pending replication.