ESRGAN Super-Resolution Models
An overview of my self-trained AI models for ESRGAN, the standard for image super-resolution.
The Sourcetex models are 4x upscalers trained on 3500 game textures. A lot of the dataset consists of concrete, brick and metal textures, which is why I recommend it for use with dirty/grungy textures. It works well as a general-purpose texture upscaler, but is sometimes not clean enough for soft textures.
Sourcetex DXTJPG Smooth - Trained on slightly over-sharpened images to produce smoother, but still detailed results.
Superscale was made to be the best real-world upscaling model there is - and it was a success, even beating RealSR (DF2K), the winner of the NTIRE-2020 super-resolution challenge, in most situations.
4x_NMKDSuperscale - Upscaling examples (Model for clean PNG images)
4x_NMKDSuperscale-Artisoftject - Upscaling examples (Alternative model for low-quality images)
Jaywreck (rhymes with JPEG) is a model made to remove image compression artifacts without compromises. It was trained on images with an extremely low JPEG quality (between 5 and 15), but the model also works for less compressed images, like JPEG quality 50 or 80.
While it was trained only on JPEG artifacts, it should work with most compression artifacts, and has been successfully tested on WEBP and MP4 (H264) compressed images as well.
1x_NMKDJaywreck - Artifact Removal examples (190k iterations)