Efficient transformer-based models have made remarkable progress in image super-resolution (SR). Most of these works mainly design elaborate structures to accelerate the inference of the transformer, where all feature tokens are propagated equally. However, they ignore the underlying characteristic of image content, i.e., various image regions have distinct restoration difficulties, especially for large images (2K-8K), failing to achieve adaptive inference. In this work, we propose an adaptive token sparsification transformer (AdaFormer) to speed up the model inference for image SR. Specifically, a texture-relevant sparse attention block with parallel global and local branches is introduced, aiming to integrate informative tokens from the global view instead of only in fixed local windows. Then, an early-exit strategy is designed to progressively halt tokens according to the token importance. To estimate the plausibility of each token, we adopt a lightweight confidence estimator, which is constrained by an uncertainty-guided loss to obtain a binary halting mask about the tokens. Experiments on large images have illustrated that our proposal reduces nearly 90% latency against SwinIR on Test8K, while maintaining a comparable performance.

The rising of deep learning has facilitated the development of single image super-resolution (SISR). However, the growing burdensome model complexity and memory occupation severely hinder its practical deployments on resource-limited devices. In this paper, we propose a novel joint-distillation (JDSR) framework to boost the representation of various off-the-shelf lightweight SR models. The framework includes two stages: the superior LR generation and the joint-distillation learning. The superior LR is obtained from the HR image itself. With less than $300$K parameters, the peer network using superior LR as input can achieve comparable SR performance with large models, e.g., RCAN, with 15M parameters, which enables it as the input of peer network to save the training expense. The joint-distillation learning consists of internal self-distillation and external mutual learning. The internal self-distillation aims to achieve model self-boosting by transferring the knowledge from the deeper SR output to the shallower one. Specifically, each intermediate SR output is supervised by the HR image and the soft label from subsequent deeper outputs. To shrink the capacity gap between shallow and deep layers, a soft label generator is designed in a progressive backward fusion way with meta-learning for adaptive weight fine-tuning. The external mutual learning focuses on obtaining interaction information from a peer network in the process. Moreover, a curriculum learning strategy and a performance gap threshold are introduced for balancing the convergence rate of the original SR model and its peer network. Comprehensive experiments on benchmark datasets demonstrate that our proposal improves the performance of recent lightweight SR models by a large margin, with the same model architecture and inference expense.