I have just received an update from DxO about NVIDIA GPU RAM processing and instability with AI masks.
This is very encouraging!
Quoted email below:
We are contacting you as part of a global communication concerning the incident affecting some users equipped with NVIDIA RTX graphics cards on Windows 11 when using AI features in DxO PhotoLab, particularly AI Masks and DeepPRIME / DeepPRIME XD demosaicing and denoising.
Over the past months, our teams have been investigating instability reports occurring during intensive AI processing. Our analysis has shown that the issue is related to how certain NVIDIA drivers handle video memory (VRAM) allocation during AI workloads when using Microsoftâs WinML framework.
Under heavy AI load, such as when combining AI Masks with DeepPRIME processing, the NVIDIA driver may not manage GPU memory allocation optimally. In these situations, Windows can interrupt the GPU task to prevent the system from freezing, which may lead to export errors or application instability.
NVIDIA has been aware of this behavior and released an initial driver update last October intended to improve the situation. However, results have been mixed depending on hardware configurations, and NVIDIA is still actively working on further improvements to their drivers.
At the same time, DxO Labs R&D has not remained passive. Over the past few days our engineers have been testing an internal corrective adjustment designed to better handle these driver behaviors during intensive AI processing.
The good news is that these tests have shown very encouraging results.
An official PhotoLab update including this corrective improvement will be deployed on March 17, and it should significantly improve stability for users affected by this issue.
At this stage, there is no action required from you.
Fingers crossed whatever is rolled out (from DxO or in the future from nVidia) further improves the situation.
Iâm still not clear why competing products (e.g. LR, C1, TopazLabs etc.) donât seem to encounter such catastrophic failure (e.g. hard crashes or a complete failure to complete the task) where PhotoLab does struggle with that outcome.
Ik think that DXO has designed its software to use as much as possible of your computing resources. Thereâs no other software that is able to instantly render the heavy AI task as DXO can.
Other software have options to render some parts manually to see the result, before exporting.
In the end, they were maybe a little bit over ambitious. But hopefully, there will come a fix. Currently, I can use most of the functions, except the predefined AI masks.
Iâm not sure I can agree, as PhotoLab is far from instant for me (at times it locks up for 30-60 seconds while rendering a preview, in fact Iâve found it quicker to actually export the shot and look at that as a preview, rather than waiting for the preview to catch up!)
Iâve found the above competition far quicker to render results.
In fairness perhaps theyâre not 100% full fidelity âfinalâ results, but theyâre more than enough to edit with and they appear in seconds, not locking up the program or the computer.
If what youâre saying is true and PhotoLab is essentially exporting the image each time you make an update using masks (or anything else for that matter) then Iâd very much like to turn that off and settle for a lower-fidelity preview while editing in exchange for rapid responsiveness and no lock-ups. (And I do already have âhigh quality previewâ and DeepPRIME renderingâ turned off in the Viewer quality/Display options tab. As far as I know, Iâm asking it to do as little as possible, and it still hangs.
I wonder what they mean by that. Didnât get any mail about it, nothing similar in FAQ.
If this was a response to an opened case, which card was mentioned?
You mentioned recently having laptop RTX5070, so the âknown WinML issueâ mentioned in PL FAQ probably affects not only RTX-x060 series, something they wouldnât elaborate on when asked in this forum.
Thanks for update, btw.
Probably different âplatformâ used, heavily redacted or frozen (C1) forums, different execution timing (there seems to be some race condition involved), people getting used, etc. Very few people would understand full technical answer
Thatâs probably due to mentioned VRAM leak, causing Windows WDDM layer to switch to using âshared GPU memoryâ (this is fully managed by Microsoft, not NVIDIA, unlike VRAM), which is much slower. I get this sometimes when working with AI masks and exports, so if e.g. VRAM usage is 11.6/12GB and âshared GPU memoryâ at 6GB or more (out of 16GB), then I restart PL. The pattern is not reproducible, so thatâs why I think some race condition is involved. Check also swap configuration and revert to defaults, if you made some manual tuning (done by many in much older Win versions) â thatâs another common reason in Win11 for performance problems.
A lot can happen before March 17, e.g. some NVIDIA updates, or last moment unwanted side-effects encountered. Common practice shows itâs better not to rely on DxO announced dates, something valid also for any other sales people.
Yes, this is a response from an open case. You are also correct about my laptop and GPU.
I have a fix for my issues which is forcing NVIDIA to be used for PL by a setting in the NVIDIA control panel. I guess this prevents windows from interrupting PL.
Do you mean âCUDA - GPUsâ or âOpenGL rendering GPUâ or both (in âManage 3D settingsâ for dxo.photolab.exe)? In NVIDIA App I have âCUDA - GPUsâ set to RTX4070 globally and âuse global settingsâ for PL9. NVIDIA Control Panel shows it a bit differently (âAllâ, with RTX4070 the only choice, marked), which is confusing. My desktop CPU does not have graphic unit (i7-14700KF).
DxO use ONNX (Open Neural Network Exchange) - what is good, as its provide standard for whatever the GPU/NPU/CPU come up (via âProvidersâ - i write about in here and here). And others use some different than ONNX (im not aware in theirs details). Note: sometimes others also has issues with drivers, like in the past Intel ARC 140 series vs. Adobe Camera Raw (ACR).
Keith, Iâm curious which drivers you are currently running and can you please clarify the AI instability to specifics? Iâm running a 1080ti latest nvidia studio drivers (581.57) and running into an issue with exporting: âInternal error (Estimation target finished in error, estimation target name: SemanticMasksOutputParametersCorrection@node_228[0].Estimation.Target0)
It doesnât matter if I choose 1080ti or CPU only, same error.
Sure, but it may take time, since my PL issues are somewhat non-reproducible.
I have 591.74 on RTX-4070 and this driver has other issues, getting card resets (with 2-3 sec black screen) just opening some Web page or using Photo app. Looks like a common problem. It doesnât happen too often and I have to live with it anywayâŚ
Thatâs what I assumed here as well. â My old built-to-custom desktop is equipped with this motherboard, has DVD and HDMI ports, but since I never used them, I even donât know if they work.
I never intended to use it for gaming, and opted for a quiet, air-cooled PC with an energy-efficient Nvidia GTX 1060 6GB. Last year it was replaced by an Nvidia RTX 4070 12GB and a more powerful power supply, long before the switch to Windows 11.
Please let me know if this setting resolves the issue for you guys.
As far as I remember, I always had this high-performance setting for DXO PL. The problems arose with the early version(s) of PL9 and the associated Nvidia drivers. Since PL 9.5 and the Nvidia Studio driver 591.74, it has been running much more reliably.
Do you mean âPreview Update (KB5077241) (26200.7922)â?
Currently Iâm at â2026-02 Security Update (KB5077181) (26200.7840)â .
I never install Microsoft âPreview Updatesâ, since Iâve lost enough mental health due to Microsoft, which they wonât refund
My drivers for the 3060(12GB) are shown in the snapshot above, 595.71, and I am still having problems with the St Pancras image on both the 3060(12GB) and even worse problems on the 5060Ti(16GB), with both machines running PL9.5.0 on Win 10.
