Different Neural Networks

There are a lot of different neural networks, in fact there are new ones every day. But here is a small list of important ones:

Relevant Networks

DAT

Transformer Architecture that is relatively lightweight. Considered a very good modern options.

DITN

Transformer Architecture that is relatively lightweight.

OmniSR

Transformer Architecture that is relatively lightweight.

SRFormer

Transformer Architecture that is relatively lightweight.

HAT

Transformer based architecture. Especially good at handling text.

SwinIR

Transformer based architecture.

compact / SRVGGNetCompact

Lightning fast architecture great for real time upscaling of videos if your GPU is fast enough.

ESRGAN

ESRGAN is what most of us started with when our community was founded. There are a ton of models for basically every task you can imagine, andz on modern hardware, it also runs quite fast. You can check the Model Database or our Discord server for model releases. The above model architectures are now considered superior, and ESRGAN is only included here due to the sheer number of specialised models available for it.

Real-ESRGAN

Not really an architecture, but a set of augmentations applied during training to make the model more capable of handling various artifacts and degradations. With NeoSR, it can be applied to pretty much any architecture. The repo contains models trained with it based on the ESRGAN and compact architecture.

Considered out of date

SFTGAN

SFTGAN is from the same developers as ESRGAN. It works by segmenting an image before it upscales them.

waifu2x

waifu2x, with GUI quite old at this point, all the models above can deliver much higher quality outputs.