Discrepancy between PiGDM reproduction results compared to original paper

[man-sean]

I tried to reproduce the results of fig.16 in the appendix of PiGDM using the provided implementation. I see a noticeable difference in perceptual quality between them for the ILSVRC2012_val_00000293 image with sr4:

Image from figure ("PiGDM (100 steps)"):

Reproduced image:

(note the sharper details in the paper's version, albeit being resized to 176x176 in the PDF)

To generate the image I used the following command:

python -u main.py \
  exp.overwrite=True exp.name=pigdm-sr \
  algo=pgdm algo.awd=True algo.deg=sr4 algo.eta=1.0 algo.sigma_y=0.0 \
  exp.num_steps=100 loader.batch_size=1 exp.seed=0 loader=imagenet256_ddrmpp \
  exp.save_ori=True exp.save_deg=True exp.smoke_test=True

As PiGDM has no official implementation and this is probably the closest to one we have, do you have any idea from where this discrepancy comes from?

[man-sean]

It seems part of the difference comes from the fact that the SR experiments of PiGDM were initialized from y noised to timestep 500 (Sec B.1 of PiGDM). This can be configured using exp.start_step=500. Now we get: