ThinkSR uses 'thought ticks' to boost image super-resolution by 2 dB

Traditional super-resolution models process images in a single forward pass, missing the opportunity for iterative refinement. ThinkSR, introduced by Zekai Shi, breaks this convention by adding a temporal dimension: the model runs over multiple 'thought ticks' (T=0 to T=4) where internal neuron-level histories evolve and synchronize. This Continuous Thought Machine (CTM) operates on local windows, each producing a compact summary representation. The novel Dense Query CTM (DQ-CTM) then transforms these summaries into pixel-aligned dense queries using a low-rank, parameter-efficient mechanism. Each position within a window receives its own evolving query, while shared thought dynamics refine the entire image across ticks. The architecture avoids token pooling, keeping full spatial resolution throughout the refinement process.

Preliminary experiments on 100 images show consistent gains: PSNR-Y improved from 28.10 dB at T=0 to 30.28 dB at T=4, PSNR-RGB from 26.63 to 28.78 dB, and mean l1 error decreased from 0.0346 to 0.0235. Every single evaluated image improved between T=1 and T=4. These results confirm that sparse latent thought representations — compact window summaries — can effectively guide dense spatial reconstruction. The work opens the door for broader continuous-thought architectures in dense vision tasks like inpainting, segmentation, and video enhancement, where iterative reasoning could dramatically improve output quality.