train dataset

[kia350]

hi can you share the training dataset ? DIV8K dataset is hart to download . thanks

[IceClear]

Sry. We do not have the copyright to distribute DIV8K. You have to refer to the official website.

[kia350]

ok can you share more about the detail of making dataset how many time repeat when use ori photo to make LR ?

[IceClear]

You may refer to the config files for dataset details.

[kangxi1818]

thank you for your interesting work, I have a question about the dataset I now have the inputs and the gts, when I geting the latent and samples, how to set the parameters --init-img of sr_val_ddpm_text_T_vqganfin_old.py to get the latent and samples, and the shape of init_latent is[1,4,64,64] how to save it as .npy and how to save the samples as .png

[IceClear]

thank you for your interesting work, I have a question about the dataset I now have the inputs and the gts, when I geting the latent and samples, how to set the parameters --init-img of sr_val_ddpm_text_T_vqganfin_old.py to get the latent and samples, and the shape of init_latent is[1,4,64,64] how to save it as .npy and how to save the samples as .png

The latents can be obtained from here, and the samples can be obtained from here. For the gts, they are 512x512 patches and inputs are obtained following the RealESRGAN degradation pipeline. Use np.save and cv2.imwrite.

[kangxi1818]

Thank you for your reply, I have another question, when I run the sr_val_ddpm_text_T_vqganfin_old.py to get samples and latents, how to set the --init-img python scripts/sr_val_ddpm_text_T_vqganfin_old.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 I mean the INPUT_PATH is it the path of inputs in my dataset Thank you very much!!

[IceClear]

It is just the name of your test folder. You may refer to the code for details. It is not a secret.

[kia350]

You may refer to the config files for dataset details. config files is lack of the procedure of how to use real_esrgan make dataset. and the test_example/OST_120.png was also made by real_esrgan?

when i used real_esrgan to get the lr-hr pair, lr was very hard to distinguish ：

[IceClear]

We just followed RealESRGAN to generate data. You may adjust the pipeline following your settings.

[kia350]

We just followed RealESRGAN to generate data. You may adjust the pipeline following your settings.

Thanks