Attention-constrained diffusion-prior super-resolution for scan-free terahertz focal-plane-array imaging

Compact scan-free terahertz focal-plane arrays provide full-frame acquisition, but the measured images are constrained by diffraction blur, finite pixel integration, array nonuniformity, and weak material contrast. We report TeraDiff, a training-free super-resolution method that uses a frozen latent diffusion model as a measurement-constrained prior for THz focal-plane-array reconstruction. A forward model and a parameterized blur simulation show that fixed point-spread-function inversion becomes unstable when the effective transfer function is unknown. TeraDiff constructs attention constraints from the measured frame and applies them during denoising, with gradients applied to the latent variable rather than to model weights. Experiments were performed on a 290–330 GHz multiplier-chain platform with a 64 × 64 AlGaN/GaN focal-plane array at a single imaging frequency of 310 GHz. TeraDiff gives the highest no-reference scores among the tested methods and improves mean stripe modulation over the raw frames, while ResShift gives the largest stripe modulation. The results therefore support TeraDiff as a balanced reconstruction prior that improves edge and stripe visibility without treating optical-reference similarity or a single modulation value as the sole evidence of THz fidelity.