4xNomosWebPhoto_RealPLKSR

Link to Github Release

4xNomosWebPhoto_RealPLKSR

Scale: 4
Architecture: RealPLKSR
Architecture Option: realplksr

Author: Philip Hofmann
License: CC-BY-0.4
Purpose: Restoration
Subject: Photography
Input Type: Images
Release Date: 28.05.2024

Dataset: Nomos-v2
Dataset Size: 6000
OTF (on the fly augmentations): No
Pretrained Model: 4x_realplksr_gan_pretrain
Iterations: 404'000, 445'000
Batch Size: 12, 4
GT Size: 128, 256, 512

Description:

short: 4x RealPLKSR model for photography, trained with realistic noise, lens blur, jpg and webp re-compression.

full: My newest version of my RealWebPhoto series, this time I used the newly released Nomos-v2 dataset by musl.
I then made 12 different low resolution degraded folders, using kim's datasetdestroyer for scaling and compression, my ludvae200 model for realistic noise, and umzi's wtp_dataset_destroyer with its floating point lens blur implementation for better control (since i needed to control the lens blur strength more precisely).
I then mixed them together in a single lr folder and trained for 460'000 iters, checked the results, and upon kims suggestion of using interpolation, I tested and am releasing this interpolation between the checkpoints 404'000 and 445'000.

This model has been trained on neosr using mixup, cutmix, resizemix, cutblur, nadam, unet, multisteplr, mssim, perceptual, gan, dists, ldl, ff, color and lumaloss, and interpolated using the current chaiNNer nightly version.

This model took multiple retrainings and reworks of the dataset, until I am now satisfied enough with the quality to release this version.

For more details on the whole process see the pdf file in the attachement.

I am also attaching the 404'000, 445'000 and 460'000 checkpoints for completeness.

PS in general degradation strengths have been reduced/adjusted in comparison to my previous RealWebPhoto models

Showcase:
Slow Pics 10 Examples