GLEAN: Generative Latent Bank for Large-Factor Image Super-Resolution

Low-Resolution (b) ESRGAN (c) PULSE (d) GLEAN (ours) (e) Ground-truth Figure 1: Example of large-factor super-resolution (16×). (a) The low-resolution input (LR). (b) ESRGAN [33] trains the SR generator from scratch, which often produces artifacts and unnatural textures. (c) PULSE [26] achieves more realistic results by GAN inversion, which, however, cannot faithfully recover the structures of the ground-truth. (d) With the proposed generative latent bank, GLEAN is able to generate output that not only is close to the ground-truth, but also possesses realistic textures. (e) The ground-truth (GT).