This is the latest revision of this page; it has no approved revision.Jump to navigation Jump to search
NMKD's User Page
My public ESRGAN models (may not be added to the wiki yet): https://nmkd.de/?esrgan-list
Here's a bunch of tips (some for beginners, some for everyone) for ESRGAN/BasicSR training based on my experience training tons of models.
- Downscale your HRs to 50% using Nearest Neighbor (aka Point) scaling if they are big enough to begin with. This will make your model sharper without creating halo artifacts.
- GAN loss is not required, disabling it can speed things up and will make models more stable. The drawback is that they are less generative, so only use this for rather "easy" tasks like cartoon upscaling.
- Validation is not necessary, so if you want to save training time, just set the val_freq to a really high number and test the checkpoints yourself.