Inquiry regarding training sequence without diffjpeg

[Mer9ury]

Hello, I'm trying to train Invertible-ISP without using DiffJPEG.
When I turned off the use of DiffJPEG, I can observe the psnr of raw image was too high and raw reconstruction loss went almost 0(1e-8).
Is it any additional technique required for training w/o JPEG simulation???