From the results，seem not good!

[BoxFishLab]

No description provided.

[zhengchen1999]

Maybe you can provide relevant examples. Or provide other methods that are superior to ours.

[zelenooki87]

Yeah, photos are not restored. Seems like only bicubic upscale is done. From test single datased x4. (example)

[zhengchen1999]

It seems there's an issue. However, the pre-trained models we provide weren't trained under blind SR conditions. @Phhofm has offered some variants that might solve this problem in #3.

[Phhofm]

Yeah it makes sense for it to look this way, official pretrains are not meant to handle degradations of this level.
My variants so far would also not work too well on this specific input, 4xRealWebPhoto is meant for photographs for example.
What would work best on this specific input were a model trained on nomos_uni with otf.
I can demonstrate what I mean, the following is an output of an RGT-S model I just trained for only 8000 iterations to demonstrate what I mean, a model fully trained this way would be able to handle this specific input best: