Vision transformer for reference-based super-resolution

Reference-based Super-Resolution (RefSR) aims to restore a low-resolution image by transferring real textures from a high-resolution reference of similar content, rather than hallucinating missing details as in single-image super-resolution. Most existing RefSR methods decompose the problem into separate matching, alignment, and fusion stages, implemented as sequentially trained modules. This decomposition causes errors to accumulate across stages and imposes inductive biases that limit what the model can learn from data. Instead, we propose to handle correspondence and texture transfer jointly through a single attention operation. We propose an end-to-end transformer that operates on the concatenated sequence of input and reference tokens, with self-attention implicitly performing both matching and texture transfer. The architecture introduces binocular 2D Rotary Positional Embedding, a shared coordinate frame for input and reference patches, and a resolution-adaptive attention scaling scheme in which the logits are rescaled by the length of the processed sequence, allowing the model to operate at resolutions beyond those seen at training time. On the standard CUFED5 and WR-SR benchmarks under bicubic ×4 downsampling, our method improves over the best previously published RefSR approach by 0.33 dB Peak Signal-to-Noise Ratio (PSNR) on CUFED5 and, for the first time, surpasses strong single-image super-resolution models on WR-SR by 0.29 dB PSNR. Unlike multi-stage pipelines, the proposed model is end-to-end trainable and requires no pretrained components, showing that a single attention operation over the joint token sequence is sufficient for RefSR. The approach is broadly applicable to image restoration settings where a highresolution reference of similar content is available.