In case you're wondering how to convert the original PyTorch models provided, or generated from the official Real-ESRGAN repo(hereinafter referred to as "the official repo"), here's a how-to guide for you.
First, the official repo comes with a simple script that you can use to convert official PyTorch models to ONNX models. If your goal is to convert the model for other purposes, I recommend you to use this script to do the trick. But before you do that, you may want to read the section # Working with models with different surffixes and do some changes to the script provided byt the official repo.
If your goal is to produce models for this GUI utility, please refer to this script(referred to as "the script") that is exactly the same one that I utilized to convert the models.
I strongly recommend you to do the conversion in the Colab notebook demo provided by the official repo, as this can save you a ton of hassles.
If you're working with the script I provided, you can continue reading. If not, you're going to apply chagnes accordingly.
The script works out-of-the-box with models that have surffixes contains x4plus(so x4plus_anime_6B works too).
If you're working with a model comes with a surffix like x2plus, you're gonna change the script's line 7 to
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)Note that the scale parameter is 2 instead of 4 of the script.
In this case, there're a little bit more you're gonna change.
line 3 to
from realesrgan.archs.srvgg_arch import SRVGGNetCompactAnd of course, line 7 to
model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu')What's more, you're gonna change the upscale parameter according to the input model's surffix, e.g. animevideo-xsx4 is 4 and animevideo-xsx2 is 2.
If you still have any questions, please feel free to draft a new issue in this repo.