that s exactly what i mentioned when I said the “whow-effect” got lost after maybe 5.4.5. 
Tried your file - no problem. Showed as 612x816px.
I used Gigapixel AI mostly for downsizing because it gave (kept) more small details (if to compare with PhotoZoom Benvista). But I cannot say the same about ver. 5.8 , though it is better for upsizing (maybe). I had to return to the ver.5.7. Pity that it is not possible to have and use both versions (5.7 - for downsize, 5.8 for upsize).
Are you running the JPGS and HEICs together?
I’ve tried both separately and together. There didn’t seem to be any difference in stability.
with this version is that the images flicker in the preview panel.
I don’t recall noticing it.
Do you find converting from JPG to DNG offers many advantages?
I think it depends on what software you use to edit the images. Overall, I think there are more possibilities with a DNG. It “feels” like I have more room to work and adjust, but, I can’t say for sure. I’ve started only shooting in RAW, and so wanted to make my legacy photos consistent in the same format.
Interesting. Photos downloaded from Facebook or Instagram are 0x0 size for me, while photos taken with my phone are flawless. I encountered this bug in version 5.8 under Windows.
I tried Reduce Color Bleed and Face Refinement too before I see this problem.
Now I tried it again. I downloaded this photo from Facebook, into My Pictures folder. I opened Windows Explorer and dragged the photo into the Gigapixel window. I got 0x0 image. 100%-os zoomot használok
https://www.facebook.com/photo?fbid=2502178603181208&set=a.474684342597321
After I opened the downloaded image in a photo editor and saved it under a different name, it recognized the size without any errors. Only this is not the original image, but a distorted copy.
I see the difference between the two in MediaInfo: the original image has ColorSpace value.
Interesting. I followed this link in your last post (above) and it shows 0x0px just as it did for you.
Using Picture Information Extractor I find the EXIF data has been stripped, all of it. However, a lot of data still remains and is readable.
UPDATE <<<
Here is the data PIE found in the image.
1 Like
Disable face Refinement, then its fine.
Hi Edwin! We have investigated your uploaded image. The reason is that, the face refinement is designed for small (low resolution) faces. Such faces can be commonly seen in vintage photos from scanning, crowd photos, photos from old devices, and photos that are strongly compressed. Meanwhile for faces that have more details, rendering them as textures just the same way as other parts of the image, can possibly give better results. The tooltip besides the Additional Settings mentions this.
In the implementation, for larger faces, they will be bypassed by face refinement function, making them to be processed as textures. Meanwhile, the faces in the border get be hard to say which way of rendering is better.
We’ve looked at the log files but the reason is still unclear why it happened to your machine. We cannot reproduce it with our machines. We will later ask for your kind helps to figure this out.
I reverted to version 5.7.3, where it recognizes the same image flawlessly. (I shortened the filename, in case that was the problem, but no)
I don’t know what changed between 5.7.x and 5.8 but the GPU performance on AMD tanked. Absolutely horrible now - takes forever to process a simple 2x resize. I am watching the process monitor and it never gets about 6% utilized. Please, whatever you did change it back to take advantage of latest gen AMD GPU’s (I have a 5700 XT).
EDIT: I have narrowed down the issue. Even with AMD 5700 XT selected it is using the Intel CPU to process. How can this be fixed?
I wonder where the two versions are picking up the image data? As I posted above, I, too, got 0x0px with the directly downloaded image you linked to.
Picture Information Extractor found the EXIF data has been stripped, all of it. However, a lot of data still remains and is readable. This is what PIE found:
Hi JML! Please refer to the below for the reason. By turning off the face refinement in this case should give you the similar and possibly better (depending on the model you are using) results than 5.4.5. For refining of faces of higher resolution, which is a new function from existing face refinement, we are working on this and can potentially delivered to you in the next release.
Hi! Could you help indicate the processor you were using when comparing? Do v5.7 and v5.8 have the same quality on cpu mode? We expect the two versions have very similar (if not 100% same per-pixel) quality on cpu. In previous versions before v5.8, gpu gives worse quality than cpu. But in v5.8, gpu should now give almost same quality as cpu.
Could you help send us the log files for processing an example image? We can look into this from there.
https://www.dropbox.com/request/WYeK0Hs8JqNJu8h3FfZc
LMK when you sent. Thanks!
After I opened the downloaded image in Photofiltre and saved it as JPG, I was left with this much data. Look JFIF datas! They differ in many ways. Maybe it also looks at version 5.8.0?
Hi yuan.liang - thanks for the reply!
Before doing my Nth install of GPAI (I’ve gone forwards and backwards a lot) I had a look at the images I did with 5-8-0 - I kept versions with all models (which will be in the archive I sent too) and am assuming I’ll get the same result away from faces with a re-install and face refinement off. I think the very compressed option (which it isn’t) is the only decent one for this image, although a little bit behind 5-4-5 still. If the next version is soon I’ll wait for that, if not I’ll put 5-8-0 back and play with some other images?
Yes at least for that specific image you sent to us, I tested it you should get a similar result.
Let me explain the context. For the gap between 5.4.5 and its following versions, it should be mostly caused by the implementation of GPU version of the models (I assume you are using gpu as processor and you are using a windows machine). An indication for this is, when you run on cpu as processor, different versions of software should have very similar results (Except for those models that we updated/improved in some releases). In this version v5.8, we actually bring a fix for this issue by modifying the gpu version of all the models. Now you should get very similar results in v5.8 as in v5.4 even using gpu (while gpu will definitely boost speed).
Another thing you could try is to play with the parameter settings. It can potentially bring you the desired quality.
We do not have a scheduled date for the next version (but we will highly possible to have a patch to fix minor issues), but it will definitely be more than 1 month later.
I can understand why face refinement would be skipped for higher resolution faces - that makes sense. However, the issue with the example I uploaded is that it didn’t just bypass face refinement, it also bypassed the normal texture enhancement. It detected the face area and enhanced everything except that. There’s just no situation in which this would be the desired outcome. Surely this is a problem, especially for batch processing?
