4x upscale is fine but 2x upsacle poor with my own finetuned model

[leya516]

Hello, I finetuned my own model with additional data following the README; Finally i got the ckpt, and i tested my images, but i found using "scripts/sr_val_ddpm_text_T_vqganfin_oldcanvas.py" with "--upscale 4" generate decent results, but with "--upscale 4" the quality of generated images drop sharply

input:

--upscale 4

--upsale 2

I can hardly tell the difference between samples 2x upsacle and inputs, but 4x upsacale sample seems all right, Could you please give me some hint of what happend? Thanks

[IceClear]

I am not sure about what happens but you can just downsample the 4x results to get the 2x one.