StableSR
StableSR copied to clipboard
IceClear
CFW module weight didn't work as expected.
Hello, I run the Test command 3
python scripts/sr_val_ddpm_text_T_vqganfin_oldcanvas_tile.py --config configs/stableSRNew/v2-finetune_text_T_512.yaml --ckpt CKPT_PATH --vqgan_ckpt VQGANCKPT_PATH --init-img INPUT_PATH --outdir OUT_DIR --ddpm_steps 200 --dec_w 0.5 --colorfix_type adain
to inference a img, when I tune the weight dec_w in the set {0.2, 0.5, 0.8} but got almost the same result, which I think lacks fidelity, how could I get SR result more similar to the original image as the paper claimed.
Here is my result:
top_left is ori image, top_right is sr result of dec_w=0.2, button_left is sr result of dec_w=0.5, button_right is sr result of dec_w=0.8
Hi. I think your results are reasonable. StableSR sometimes may not show 100% fidelity given the powerful generation capability of the fixed Stable Diffusion. And I am also not sure if other methods like Real-ESRGAN and LDM could meet your needs. Since SR needs to generate new details according to the input and sometimes it may lead to some differences. BTW, I will appreciate if you could provide the comparison between your results and the results you refer to in the paper. I do not see an obvious gap between them.
Hi, it seems it's the problem of sr scale, when I do 1XSR, the result lacks fidelity, but 4XSR result is much better.
left is ori img(720720), mid is 1XSR result(720720), right is 4XSR result(2880*2880)
4XSR result is almost the same as the result you provided. Why is 1XSR not so good?
Hi. I think it is because 4x enables more generalization while 1x provides more limits on the results.
好，看来是sr比例的问题，当我做1XSR时，结果缺乏保真度，但4XSR结果要好的分割。左边是ori img(720 720)，中是1XSR结果(720 720)，右边是4XSR结果(2880*2880) 4XSR结果与您提供的结果几乎相同。为什么1XSR不那么好？
Hello, may I ask how to change the code to be able to 4XSR?
好，看来是sr比例的问题，当我做1XSR时，结果缺乏保真度，但4XSR结果要好的分割。左边是ori img(720 720)，中是1XSR结果(720 720)，右边是4XSR结果(2880*2880) 4XSR结果与您提供的结果几乎相同。为什么1XSR不那么好？
Hello, may I ask how to change the code to be able to 4XSR?
In the python file : scripts/sr_val_ddpm_text_T_vqganfin_oldcanvas_tile.py, there is an argument: parser.add_argument( "--upscale", type=float, default=4.0, help="upsample scale", )