Open xsacha opened 4 years ago
The model will have to be retrained. I am not sure if the quality of images is good, but the model generates x8 upscaled images (256x256) since the network is composed of fully convolutional layers.
The model will have to be retrained. I am not sure if the quality of images is good, but the model generates x8 upscaled images (256x256) since the network is composed of fully convolutional layers.
Hello,thank you for your work.But i got quite bad results.I put pictures(128128 bicubiced by 1616 ) in CelebA to test.Can you tell me if i am right or not?Thank you.
Hi, you should make the input image (16x16) from the HR image (128 x 128) with the progressively downsample. 128x128 -> 64x64 -> 32x32 -> 16x16. We found that directly downsampled image is slightly different to progressively downsampled one.
Hi, you should make the input image (16x16) from the HR image (128 x 128) with the progressively downsample. 128x128 -> 64x64 -> 32x32 -> 16x16. We found that directly downsampled image is slightly different to progressively downsampled one.
Thank you very much.And the input size should be directly unsampled to 128 or progressively upsampled to 128?It's kind of you.Thank you very much.
Uploaded trained model can generate directly upsampled 128x128 image.
Uploaded trained model can generate directly upsampled 128x128 image.
Hello,thank you for your reply.I tried as you told me.But i got worse results.
Could you send me your test image? I will try it. Since this model was trained with bilinear downsampled images, however, the natural low-resolution image may not be upsampled successfully by this model.
It's one of my test picture.And the dataset i use is also celeba,same input and output size.
------------------ 原始邮件 ------------------ 发件人: "notifications"<notifications@github.com>; 发送时间: 2020年1月9日(星期四) 晚上9:23 收件人: "DeokyunKim/Progressive-Face-Super-Resolution"<Progressive-Face-Super-Resolution@noreply.github.com>; 抄送: "泱渶"<2676426668@qq.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [DeokyunKim/Progressive-Face-Super-Resolution] Input Size (#8)
Could you send me your test image? I will try it. Since this model was trained with bilinear downsampled images, however, the natural low-resolution image may not be upsampled successfully by this model.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
I can not find sample images.
You can try pictures starts from 801. jpg in celeba,I downsampled pictures to 16 slowly.
---Original--- From: "DeokyunKim"<notifications@github.com> Date: Thu, Jan 9, 2020 23:13 PM To: "DeokyunKim/Progressive-Face-Super-Resolution"<Progressive-Face-Super-Resolution@noreply.github.com>; Cc: "Comment"<comment@noreply.github.com>;"yangyingni"<2676426668@qq.com>; Subject: Re: [DeokyunKim/Progressive-Face-Super-Resolution] Input Size (#8)
I can not find sample images.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
@yangyingni i test with celeba imags of 178*218, got bad results,too. Do you solve it?
Does the input size have to be 16x16? What if we want to use custom size, or even 32x32?