How to generate 1X SR without zooming the LR ?

[ManuBN786]

I would like to use the DoSSR for microscopicSR using BioSR dataset.

In BioSR both the LR & HR patches are (256x256).

So I would like to use the pretrained models to get the 1X SR results.

But the SR results I get for 1X are zoomed & cropped even if I use the --disable_preprocess_model.

LR test image (256x256)

SR prediction (256x256)

The command is : 'python scripts/inference.py --input test_512/ --config configs/model/cldm_v21.yaml --ckpt pretrained/dossr_onestep.ckpt --steps 1 --sr_scale 1 --color_fix_type wavelet --output predictions/ --device cuda --disable_preprocess_model'

[lyxlynn]

In fact , we only train our model for 4x task. But you can have a try to use 4x model to test your image. Just downsample 4x to your LR image and then upsale 4x to 256*256 , and then change your command "--sr_scale =1" to "--sr_scale=4".

[ManuBN786]

My task is 1X only.

So I will train my own model for 1X SR.

I will post some questions if I face some issues in training 1X SR

Thanks for the fast reply