Sure thing, here’s what I was testing with: https://user.fm/files/v2-d507aee33613b94e9eff0bede6cec75a/RAW%20and%20DOP.zip
Don’t mind the absolutely horrendous edits, I just fooled around aimlessly for purposes of testing!
Oh, also I’m on a Mac, and I know the DOPs aren’t completely interoperable between Mac/Win. So hopefully most of it works for you all.
It is pretty pointless sending random failing RAWs and DOPs unless we have some post processing imaging experts who worked on PL9 here?
If you think about it, it should be DxO asking for the RAW. I mean, they might learn something from it. The fact they are not asking is interesting………… It’s almost like they knew that their minimum computer requirements were rubbish?
Yeah it’s pointless in terms of any effective means to actually address the issue.
Only serves to satisfy some people’s curiosity, really. And wasn’t difficult to upload, so I figured “hey, why not.”
I can only assume they scrambled to put these AI features into their application with not enough resources to get it done right. Having minimum requirements any higher than what they already have suggested would be–already is–ridiculous to ask of customers. They obviously don’t want to admit that they under-cut them, since they already were asking for so much power, and are probably hoping/praying that they can fix the app before too much customer-retention damage is done. Capture One and Lightroom can do all the same AI masking (and better) on much less powerful computers.
DxO’s history is quite interesting. I thought that they had left their previous problems behind them but maybe not. It’s a shame.
When the business was split and venture capital brought in, well, priorities changed. I’d like to think that DxO are reading the numerous threads here and elsewhere but I think they are well aware of the various issues, and either have no intention of solving them or not the resources to do so.
I will check the X-Trans DP against the DP in PL7 and if I don’t see a significant improvement, I’ll just revert to PL7. I suspect that I will keep both for many years. The AI masking means nothing much to me and I have survived without it , so no great loss. The loupe issues are a pain, which is where PL7 comes in.
On Reddit and some other sites there appears to be a plethora of people who run old copies and have no intention of upgrading. Unfortunately I think I might be at that point now.
I won’t be renting off the likes of Adobe or anyone else. I suspect PL7 and PL9 will run for a very long time and unless you are printing large photographs, you’ll barely notice any difference on a PC screen.
Many thanks for the image and the dop!
I made a few measurement, and has some observations. May some of it interesting.
I’m under Win11. Its read Mac .dop fine. In any change on values → its can’t write the changes to dop well → so PL crash. However enable/disable masks works (i guess its not save this instantly). Update: No, its not.
It has 5 mask (from bottom to top):
I made a copy of the RAW, and copy / paste all correction setting from yours to get ‘clear situation on Dop’. But i also test with yours (no difference)
I switch to CPU only mode, to get a bit ‘larger duration’ for more better measures.
I measure the ‘Full preview in progress’ duration.
Before each test: i select the photo (i made with Virtual copies), wait until ‘Full preview’ finish → Quit from PL → Start PL → And the last selected is open by default → does measurements. So, each of the test runs is ‘clean state’..
Here the results:
Rows:
Row 8 → its a ‘baseline value’, as all mask disabled → that’s the value what the photo (with global corrections) is loaded (Full preview in progress)
Column: ‘Unit - Baseline’ → its the mask(s) calculation value.
Observations - part 1:
The tested AI pre-defined masks runs in the same ‘duration’. Doesn’t matter its People or Clothes. All like 43 ‘unit’
The AI manual (selection) mask ‘duration’ is very small, and seems more Manual AI mask not change too much overall.
Observations - part 2:
The matting (combination) of masks also takes time (not surprising)
As if all mask enabled → 210 unit. The sum of the separated mask values → 137.
So, mask ‘matting’ takes like 35%
Observations - part 3:
The AI manual vs AI pre-defined mask calculation value looks like 20%. So, 1 (one) AI pre-defined mask takes like 5 times more than one or two (1-2) Manual AI mask.
Observations - part 4:
I re-create masks from scratch, where the manual AI masks is very-very same (of course pre-defined masks is the same) - but not the same changes on mask. When all mask enabled, the result is the same: 262 unit. In few cases it was more, but after multiply runs i think its the same. The baseline value is also the same.
Remarks:
Overall:
Paris is beautiful in each season. And when i open, i see instantly where the masks!
But goes back to the ‘original point’:
At general i think its mainly because of ‘pre-defined’ masks. And Export is a separated process. In Win its the DopCor process. And as its a separated process, its ‘calculate again’ the masks.
May export has its own issues (i read in the forum), but i think most of it related with GPU VRAM (or whatever its called in Mac) amount and may some constellation of things.
I think, near all AI mask ‘issues’ (performance or ‘internal error’) goes back to ‘pre-defined’ masks. And mainly ‘pre-defined’ mask vs GPU VRAM amount. And this may also impact - in some case - the export stuffs. As export use the same masking method!
@Fineus - its may interesting for you.
I not see this pointless. I think its not ‘random’. It has a reason(s) behind it. Like patterns, limits, etc. May for some its still looks like ‘random’, but may for someone its helps to see the reasons behind it.
I do a lot of measurement, observations, and comment a lot about that, like GPU VRAM usage calculations, and lot of things. I even work out (with other forum colleagues) how PL9 can work with only 4GB GPU VRAM with Manual AI mask + DP3 export. So, i pretty sure its not ‘random’.
And also good to see the masks, etc. Seeing is believing. And may you can test in your computer. And may can ensure others: its works for me.
So basically you might be able to render an AI mask as long as it’s not one of the predefined ones? Although I have never used the predefined AI masks and it still failed to export.
It’s not random because it happens every time, kind of the opposite of random. In fact it’s very reliable.
In short, even if you have above the minimum hardware stated by DxO, you still might not be able to use al the new shiny toys of PL9. Which is exactly what we were saying from the beginning.
Could you let DxO know please?
The main reason that it’s pointless is that unless you can replicate the exact system that was being used then nothing matches.
It’s like going out one day, taking a photo. Then going out the next day and taking the same photo but they don’t look the same. That’s because it’s a different day, different light, there are variables.
But it’s good to see your conclusion is basically what we believed all along.
But curiosity may can lead to discoveries? Isn’t it?
Christopher Columbus may ask: What’s in the other side of the sea? India? America? ‘Here be dragons’? Or we fall from the edge of the world? Lets go and see it. 
Discoveries in this forum already lead to some other discoveries.
And some of this discoveries may lead to better understanding what s happen, and may even lead to ‘spare money’! Yep! Curiosity may lead to spare bucks (or Eur, etc.)!
In the forum we find out how its can work with only 4GB GPU VRAM (and what’s is the ‘max’ can do, need some basic workaround) → If someone has 4GB GPU, and read it in this forum → and if try it and if works for him → and if its just okay for him → may not buy new notebook (with more GPU memory) / new GPU. I definitely don’t have any budget for new GPU (i have ancient 4GB)-> So i do it until i found / see some solution what was acceptable for me, like a trade-off. This trade-off is Okay for me, but may for some not. So, may this discoveries may helps for users to ‘max out’ what currently (!) they can do/use with their GPU VRAM amount.
The ‘actually address the issue’ is hard to say. As i think, its not even easy to define ‘what is the issue’.
PL 9.2.1 release note say like (Windows):
Minimum system configuration: … For DeepPRIME 3, DeepPRIME XD3 X-Trans, and AI Mask: NVIDIA RTX™ with 6GB of VRAM with latest drivers AMD Radeon RX6000 series with 6GB of VRAM with latest drivers"
Its just say DP3 and ‘AI Mask’.
Its not say DP XD2s just DP3, its not say ‘Pre-defined AI masks’ just ‘AI mask’.
Its also not say anything about Loupe, DeeprimeRendering, Paralel export count and so on. All of it use a bit more/additional GPU VRAM → like DP XD2s seems use more then DP3 and so on. All in the ‘top of it’ over DP3 and Manual AI mask use some more GPU VRAM → And may 6GB not enough in the end of the day.
I think with 6GB GPU its can work fine if NR DP3 only and AI masks is ‘Manual’ only, and no other apps running, etc. It’s may in the very edge, but i think its can work. Of course driver quality (seems in nVidia side) may vary.
And wasn’t difficult to upload, so I figured “hey, why not.”
That’s the Spirit! Thanks a lot! 
I not able to describe exactly what behind the export time you observe. I try. May export has some issues in his own, but i think its connected with pre-defined masks. May worth to try to turn off DeeprimeRendering to ‘shave down’ some GPU usage. May.
But its lead other observations like: seems AI pre-defined masks take a same time to calculate and doesn’t matter what the AI pre-defined type (at least in the tested masks) - what is ‘in line’ with my previous observations with Manual AI masks. Masks ‘matting’ also takes time like 1/3 of all. And Mac .dop works fine in Win (but only in read mode). 3 thing, what frankly i’m not read in this forum. Some of obvious like Mask ‘matting’ takes time - but now at least we have some idea about this ‘time’.
Doesn’t work here on Windows. If PL doesn’t crash straight away, I can click on the three first AI mask (People, background, clothes) but neither see their selections nor their effects. – Trying to examine further crashed …
Hey it was good for you to take a look. I am sure everyone appreciates the time and effort.
Shaving down the GPU usage won’t help. My GPU never maxes out and if it did it can delve in 48gb of DDR4 RAM. But it didn’t.
If I was a betting man, I suspect that the problem is connectivity. Everyone fixates on the core components. RAM, VRAM, GPU and CPU. But what is often missed is the connectivity between the software and these components.
Think of this like driving from Lands End to John O’Groats (for non-UK - a long way. 837 miles by road, 14 hours & 30 mins). You start off, the road is a country road, you have to start slowly, it takes a while to get to a decent A road (dual carriageway) then you can get going. Sooner or later you have to stop for fuel, the petrol station has fuel (great news) fill up and continue driving.
The roads are the connection between the petrol stations (Pc components). The petrol stations will always (almost always) have fuel. So the components can deliver but how slow are the roads? If the connections in your PC cannot handle the volume of information then they will back up and time out.
So the problem could be motherboard related?
Best comment ever!
First of all, the photo is less important, the DOP is more important. But its better to use the dop with the ‘original’ photo. for sure.
Sometimes the ‘check’ is not really about to ‘pinpoint’ (replicate) the problem, but more to ‘narrow down’ the issue. Like: PL version is latest? → In ‘CPU only’ mode its works? → GPU driver versions already tried? → In ‘GPU mode’ If only one (1) photo in the folder, it has some editing, DP3 NR, and 1 (one) non AI mask in there (like CP), and after PL restarted, its works? → Repeat the previous, but with one (1) Manual (Selection) AI mask, its works? And so on. Like ‘raise the bar’ in each steps.
@unchdxoly Thanks for that! My first attempt to open the first AI mask produced
But after a restart PL9.2.1 found the image and apart from the delay with the original rendering, the loupe changes with my 5900x 50506TI(16GB) were instantaneous.
Export with XD3 was 15 seconds (image and export located on an NVME).
A batch of 30 VCs took 2 minutes 12 seconds to export, 4.4 seconds per copy.
Rendering one of the VCs that had never been viewed before at 1600% took 7.544 seconds, according to my stopwatch.
@Wolfgang I skimmed (skipped) over your post and when I just went back to review the masks only the first two show in the display as you stated.
Apart from the initial Dump + fail I haven’t had any other problems while investigating the image @unchdxoly, until I went to review one of the VCs and
VRAM usage is nothing extraordinary!
I’m just a muggle and am allergic to spreadsheets, but I couldn’t resist throwing that RAW file at my setup to see how it performs. I was not trying to duplicate the full testing here, just curious at what results I got. If it adds some insight to this conversation then that’s great. If not, ignore me. Please don’t try to drag me into the PC/Mac religious war - this is my setup and I’m happy with it. I can’t wait to jump onto an M5 mini or Studio though lol.
Mac mini M4
macOS Sequoia 15.6.1
32gb RAM
2Tb SSD
All images , DOP etc on external Samsung T9 2Tb SSD
Photos, Photomator, PL, Vuescan, Safari with about 100 open tabs (ya I know) and about a dozen other apps open.
Photolab 9.2.0 build 31
Viewer quality: Enable high quality previews / Enable DeepPRIME
XD-3 enabled
I randomly threw 13 mostly AI-chosen masks on a copy of France-20250911-0007 (the version with the DOP opened fine BTW) slammed the controls about, threw a bunch of global corrections on to boot.
At no time did the Correction Preview take more than ½ second, the Loupe at all sizes was near-instantaneous.
Exporting a full res full size JPEG took approximately 11 seconds.
I usually edit with Deep Prime previews OFF, as performance is a bit slower with my Z8 raws.
YMMV
I tried it twice (on different days) and each time I got different results, now including crashes.
While I was able to see the mask of “AI Add a selection” and/or “Add an area”, none of the predefined ones did show up. … Also before, I couldn’t open/read dop files from the PL 9 Mac version.
@Wolfgang I thought it was going quite well (apart from the initial PL dump) until I properly read your post and then things started to unravel somewhat!
Every time I try to delete one of the AI masks PL9.2.1 Dumps, I should have submitted every one to DxO to make a point!!??
Back to coding it is ultimately more useful and certainly a lot less frustrating!!
@Wolfgang An update, sort of!!??
I installed one of the 591.44 variants available, but on my 5060Ti(16GB) I have problems with the NVidia drivers rejecting an update attempt because it states that the current driver is newer or the same as the one I am trying to install.
So I installed a much earlier driver and then installed the 591.44 driver successfully!?
When I tried to open the forum in the Opera browser I got a white screen with two solid black boxes! A restart cured that?
Initial opening of the test image still took a while (17 seconds) and the first attempt at deleting the ineffective masks caused a dump and a restart but I think I attempted that before the restart.
So the deletion was finally successful and I added three more masks, copied the adjustments and attempted to select all but the [M]aster and that resulted in another Dump and Restart.
This time I selected all the copies, deselected the first [M]aster image and deleted all the VCs and made a set of 29 new VCs all with the new complement of AI masks.
The test export was successful
This is the [M]aster with the “new” AI masks and 29 VCs, 30 images in total from a Win 10 system. @calvingrinder you might like to see if your Mac will accept Windows DOPs with AI masks. As before my images are on an NVME rated to be 4/5 MBs, 4MBs write and 5MBs Read, and the exports are going back to that same drive.
Please note that the 415.5KB DOP contains the [M]aster and 29 VCs while the 13.9KB DOP contains just the [M]aster.
France-20250911-0007.RAF (21.3 MB)
France-20250911-0007.RAF.dop (415.5 KB)
France-20250911-0007.xmp (917 Bytes)
This is a DOP with just the [M]aster image
France-20250911-0007.RAF.dop (13.9 KB)
@unchdxoly Thank you for the image, @Gareth while I feel that the latest NVidia drivers are making some improvement, the stability is still not what it needs to be, because I am using a 5900X the restarts of PL9.2.1 are not as painful as they might be but it is still tedious to have to keep restarting PL9.2.1 after a Dump and all future Dumps will be forwarded to DxO.
I am sure it’s on DxO’s “to do list” 
@Gareth Hmm!?
It is but how soon will it actually happen?
DxO will be due to start their PL10 Beta testing sometime towards the end of January and we are not really out of PL9 Beta Testing yet 
!
