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

SKILD unifies image generation and super-resolution in a single unconditional diffusion model by treating scale as an explicit coordinate in the forward process, attenuating content from fine to coarse while injecting spectrum-matched Gaussian noise in k-space. At inference, the same trained model handles both tasks by varying only the starting scale, enabling continuous upsampling at arbitrary magnification factors without task-specific retraining.