Comparative results (extra)
Our arbitrary-scale upsampler produces the sharpest images at reduced computational costs (lowest memory and number of operations). The presented results illustrate its performance when combined with different encoders and at different upsampling scales.
Encoder SWINIR, scale 4x, dataset: Urban100, image id: #008.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right). |
Encoder EDSR-baseline, scale 4x, dataset: Div2k, image id: #0826.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right). |
Encoder RDN, scale 4x, dataset: Div2k, image id: #0891.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right). |
Encoder EDSR-baseline, scale 3x, dataset: B100, image id: #86016.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right) |
Encoder EDSR-baseline, scale 4x, dataset: Urban100, image id: #030.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right). |
Encoder RDN, scale 8x, dataset: Div2k, image id: #0853.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right). |
Encoder SWINIR, scale 4x, dataset: DIV2k, image id: #0831.
| LR input | CUF (left) vs. LIIF(right) | CUF (left) vs. LTE(right). |
Acknowledgements
The website template was adapted from Ref-NeRF webpage with authors consent.
The ablations replicating previous models were made under LIIF, LTE, SwinIR and ABPN public codebases. We thank the authors for sharing their codes.
Please refresh the page if videos are not loaded or appear unsynced.