Upscayl is a front end for Real-ESRGAN
What Real-ESRGAN is good at
Real-ESRGAN (Wang et al., 2021) was trained on synthetic real-world damage — blur, noise, resizing and JPEG, applied twice — so it handles phone photos, screenshots and reposts well [3]. It is open source under BSD-3-Clause, with a separate lighter model for anime [4]. Kenerate's default engine is Real-ESRGAN, so on the same image the two give similar results; the difference is where it runs and what else is available.
When a different engine is better
Real-ESRGAN is a GAN model tuned for sharpness. For faces in old photos a restoration model is better; for art where you want new texture, a creative diffusion-style upscaler is. The perception–distortion trade-off means no single model is best at both [5].
























































