The effect of using display pictures is very poor

[learn01one]

The effect of using display pictures is very poor

Result picture

run command：

python scripts/sr_val_ddim_text_T_negativeprompt_canvas_tile.py \ --config configs/stableSRNew/v2-finetune_text_T_768v.yaml --ckpt ./models/stablesr_768v_000139.ckpt \ --vqgan_ckpt ./models/vqgan_cfw_00011.ckpt --init-img ./test --outdir ../OUT_PATH/ \ --ddim_steps 20 --dec_w 0.0 --colorfix_type wavelet --scale 7.0 \ --use_negative_prompt --upscale 4 --seed 42 --n_samples 1 --input_size 768 \ --tile_overlap 48 --ddim_eta 1.0

The experimental results show that only the image size has increased, but the quality is still very poor. Is there any problem with running the command?

[IceClear]

Hi. You are using a zoomed image, i.e., the LR image is already a 4x bicubic upsampled version of the original LR image. We guess that a too-large upsampling scale, i.e., 16x may lead to undesired results sometimes. This case is not very common and we did not do many tests on such cases. A possible reason can be the large gap between the training and inference.

[IceClear]

You may test more images using the original images here. If you are interested in finding the reasons, you can first try upsampling these images with 16x to see if such a case happens to most of the images. My current guess is that the upsampling scale is too large, not sure about it. I am busy recently and may have a test later.

[IceClear]

I tested on the original image using DDPM 50 steps and it works. I think DDIM 20 steps should work too.

The original image for your reference:

[learn01one]

Hi. You are using a zoomed image, i.e., the LR image is already a 4x bicubic upsampled version of the original LR image. We guess that a too-large upsampling scale, i.e., 16x may lead to undesired results sometimes. This case is not very common and we did not do many tests on such cases. A possible reason can be the large gap between the training and inference.

Thank you very much for your reply. I have tried other samples and it is indeed as you analyzed. By using the original image you provided, the effect is normal.