search

rlghksdbs / CASR

11 stars 1 forks source link

why pretrained_models are x4 version? #2

Closed zephyRMountain closed 3 months ago

zephyRMountain commented 3 months ago

As your youtube video shows, pixel unshuffle 3x down sampling has the best performance, and sr_demo.py also has a default weights named PlainRepConv_BlockV2_x3...

zephyRMountain commented 3 months ago

So Many Unnecessary Codes!

zephyRMountain commented 3 months ago

But it is a great work undoubtfully. Looking forward to your update!

rlghksdbs commented 3 months ago

Thank you for your interest in our work.

The competition is to ultimately upsample by a factor of 4 to create a 4K image.

In the config of sr_demo.py, it says ‘PlainRepConv_BlockV2_x3...’ because we copied the code from our 2023 CVPRW paper.

That's just the name, so you can ignore that and just paste the weight path you learnt into config.

I kept putting off cleaning up the code, so I simplified and updated only the necessary code.

I'll update details like ‘PlainRepConv_BlockV2_x3...’ in the config and the name of the definition soon.