cetini18/hr2lr-noise-sweeps

HR to LR generator: noise-realism sweeps

Pictures shown on the project page Noise-realism runs (route /research/super-resolution/lr-data-generation/hr2lr-ablations of nitec427.github.io). Each picture is a 256 x 256 low-resolution image generated from a clean high-resolution image with the degradation (blur, noise, compression) of a reference low-resolution crop. All files are lossless WebP, because the noise must not be compressed.

File names

file	content
hr<k>.webp	clean HR crop, k = 0..5 (0 RealSR Canon_049, 1 SVSR 00139_6D, 2 RealSR Canon_017, 3 RealSR Nikon_007, 4 SVSR 00136_6D, 5 SVSR 00156_6D)
ref_<level>.webp	reference LR crop of that degradation level (a real image of a different scene; one fixed crop per level)
real_<level>_hr<k>.webp	real LR of the same crop as hr<k> where the dataset has one (SVSR: every ISO level; RealSR: only the level the HR came from)
<run>_<level>_hr<k>.webp	generated LR of run <run>, content hr<k>, reference ref_<level>

Levels: realsr_{Canon,Nikon}_x{2,3,4} and svsr_iso{100,200,400,800,1600,3200,6400,12800,25600,51200,65535}. Runs: long_E0_mse (baseline), round 1 q_A5_B_head_psd, q_N6_all, q_N4_head_psd_hiiso, q_A3_head_gan02, q_A6_head_psd_160k, long_E2_psd, q_N5_head_psd_gan, q_A4_head_psd_hiiso8, q_F4_inj_psd1, q_F2_inj_psd_enc; round 2 q_G1_ngan_vec ... q_G7_ngan_desc_nobins, q_B1_head_gan_hiiso8 ... q_B4_head_psd_gan_hiiso4_from1600. Every run of a given hr<k> uses the same random noise seed. Run descriptions and metrics are on the project page.

Licence and sources

The clean and real images are crops of the SVSR dataset (Aakerberg et al., WACV 2024, CC BY-NC-SA 4.0) and of RealSR (Cai et al., ICCV 2019, research use only). The generated images come from models trained on those data. This dataset is therefore shared for non-commercial research under CC BY-NC-SA 4.0; please cite the original datasets.