Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution

Researchers introduced SKILD, a scale-invariant k-space diffusion model that treats scale as an explicit variable.

By changing only the starting timestep, one trained model can both generate images and perform 2x to 8x super-resolution.

The approach works without conditioning branches, classifier-free guidance, or retraining.

It reports strong results on CIFAR-10, ImageNet super-resolution, and critical Ising system reconstruction.