The original images used to train the new model have a resolution of 1024x1024, compared to the previous 512x512. They were taken from the IR105 layer instead of IR108, as this was the only layer available in such high resolution from the provider. I’m also working on a new model, but at this stage, only a prototype version has been trained.

Metrics

The images were compared using the following methods:
• PSNR: Measures the ratio of useful data to noise. The higher, the better — typically ranges between 16–30.
• MSE: Calculates the average squared difference between corresponding pixels. The lower, the better. However, it’s less comparable across different resolutions.
• SSIM: Indicates how visually similar two images are. Values range between 0 and 1, where 1 is a perfect match.
• LPIPS: Also measures similarity, but by dividing images into smaller patches and comparing those. The lower, the better, with a range of 0 to 1.
• FID: Unlike the others, this doesn’t compare images one-by-one but evaluates the whole set of images as a distribution. Lower scores indicate better similarity to the reference set — 0 means perfect match

Results of the old model over 3 weeks