XPixelGroup
Weird (behavior) testing HAT models.
HI ! First, thx for your work, it's awesome with most of images ! But i have a problem testing the image i work on. I work on marriage acts scans, so mostly handwritted text data on uniform backgound.
And i don't really understand HAT behavior : if i give it a low-quality downscaled 64x64 image from a 256x256 crop in my base high-res scans, all models works well, the bigger being the better, and GAN hallucinate a lot. This i think is all mormal.
If i give it either a real 64x64 crop, not downscaled, or the 256x256 original from the test above, the result shows absolutely no imprevements from the image i gave it. It mostly add noise and not make the text crispier or anything. The upscaled image from the 64x64 downscaled input is even miles better ! the GAN version here works really well in this case though.
If i give models another 'high quality' input which is not text though, like a crop in an anime picture, they all work perfectly.
Do you have some hint, or some explaination for me ? Thanks !
@MrBread13 The GAN version considers multiple degradation types. Thus the model can deal with many complex cases, including the bicubic degradation (the low-quality downscaled 64x64 image from 256x256) and the real unknown degradation (a real 64x64/256x256 input ). The provided Non-GAN models only consider the bicubic setting, which is trained for academic comparison. It means that these models have the strong ability to deal with the bicubic downscaled image, but are difficult to process the other degradations well. For the high-quality input you mentioned, it is possible that processing these images requires very little model capability. So both models can handle well.
