
Read the full blog entry here.
2020-02-11 14:10:01
2020-02-11 16:46:27
2020-02-12 08:44:48
Closer in what way? As far as we've seen, their approach doesn't provide real-time (or close to real-time) feedback and that's what we'd be aiming for.
2020-02-17 10:27:44
Ah I understand. Yes their approach is very far away from being realtime. That's true. But what I meant was the way new texture is generated. I don't know how they do it but maybe you have the answer.2020-02-12 08:44:48
Closer in what way? As far as we've seen, their approach doesn't provide real-time (or close to real-time) feedback and that's what we'd be aiming for.
As for 3x — that's the most computationally efficient amount and that's how much the underlying algorithm increases every image, even if you choose a smaller size in the Image Size dialog. We simply reduce the result to the required size afterward using the Bilinear algorithm so you can use the Scale Images action to scale the image down. If you choose more than 3x, the algorithm runs twice, so you can simply apply the Increase Resolution of Images action twice to get the same result (and then use the Scale Images action to get an exact size).
2020-02-17 13:55:23
We can't know for sure what kind of approach they took but it doesn't really matter all that much — we plan to keep improving our algorithms (not just for ML Denoise) regardless. And if you ever find any images where you think we could do better (either generally, or compared with other apps), sharing one or two samples with us at support@pixelmator.com helps us a lot!2020-02-17 10:27:44
Ah I understand. Yes their approach is very far away from being realtime. That's true. But what I meant was the way new texture is generated. I don't know how they do it but maybe you have the answer.
Would love to see this kind of batch processing inside of Pixelmator Pro to see an actual progress bar or something like that.
2020-02-20 13:41:06
2020-05-16 09:20:03
2020-05-18 11:42:54
Sometimes, though not usually. If the image is quite small, the amount of detail in it is constrained the number of pixels. For example, if there is a face represented by, say, 48 total pixels (8 x 6), you can't make the edges more natural without adding more pixels to the image, it's physically impossible. If the image is larger and the face is represented by, say, 192 pixels (16 x 12) but it's blocky or blurry because someone previously upsampled it using nearest neighbor/bilinear, then it would theoretically be possible to increase detail while using the same number of pixels, but that would require a different approach. In a case like that, the ML algorithm would probably need to be trained to recognize previously upsampled images and detect the resampling algorithm used in order to undo those changes and then, while keeping the same size, upsample the image again. Also, this isn't a very common problem — usually, the images are just very small, so the pixel size needs to be increased to physically be able to create natural edges and reintroduce detail.
2020-09-07 23:13:26