As per the title I find any local AI based edits can quadruple + the export time - I have a new win 11 rig with a 5070 Ti and a good i7, 32GB DDR5 RAM, so this was unexpected - if I do an export without AI local edit masks it is blazingly fast. The more AI edits I add, the greater the impact on export time. Anyone else getting this? It’s literally about 5 sec to export a Deep Prime 3 file from my Sony a9 if there are no AI based mask edits and about 2-4 minutes if there are AI based masked edits
I’m finding the same thing, 5-10 seconds for little to no AI edits and up to 3-4 minutes for more AI edits. Besides that, I’ve had a lot of other problems with error msgs, lagging, unable to export etc. . What’s weird is, today I haven’t had any of those issues (except for the longer export time on AI edited images) and I haven’t done anything different.
DXO is well aware of various AI mask performance issues. They are actively working to improve AI mask creation and export efficiency as well as fixing the internal errors many have noticed when creating AI masks. However, I have no idea when these fixes will be available or whether eliminating all the performance issues will require the implementation of multiple fixes over time.
Mark
Where did you read that they are aware of the issues and working on fixing things?
one can assume they use what they sell no ? they were not able to finish product in time and just like Fuji ( remember their “kaizen” marketing ) sell semi-cooked stuff to get money to finish the product … it will pass
I am not at liberty to discuss that. You will have to use your imagination.
Mark
2 Likes
ok, does not NDA include the part of non disclosing the existance of NDA itself ?
Yes it would. However, I made no reference to an NDA. You just assumed it.
Mark
Ok.
I just wish they could inform people on here or someplace that they were aware and working on it so that people didn’t waste hours trying to figure out what the problem is and know that a fix was coming.
DxO will not do that. The first any of us will know about it is when they release an update. Even testers don’t know when that will be.
Mark
The above discussion complements my own experiments with PL9 Elite Complete under the 30 day trial license. I have re-done some of the raw images using the same .dop file and compared the PL8 Elite Complete JPEG output (accepted by a client) to the PL9 Elite Complete JPEG output, trying to use the latest features in PL9. First, as I commented elsewhere, looking at the two different JPEG images (identical crop, PL8 vs. PL9), the differences were not that much. However, although I did not mention it in my previous comment, export to a JPEG processing time did dramatically increase. Thus, I will wait until Black Friday pricing for PL9 Elite before licensing it for fee. Although Mark will not explain how he knows that DxO is “working” on the export processing time increase issue, it is reassuring that someone on this forum (Mark in this case) is willing to post a comment that DxO software engineers are working on the issue. DxO is completely silent on how the “AI” was implemented, presumably as a neural network using the GPU presented to the application rather than a simpler expert system implementation. If indeed an AI, the issue of how the AI was trained and if it will continue to be trained from the actual images processed by an individual photographer. Moreover, if the AI is trained from the images of the photographer, does the AI “call home” and provide DxO with that information to improve the AI ability for all who use PL? If so, is the AI violating the intellectual property of the photographer who “took” the image? It would be better were Marie or another “authorised” DxO staff person replied, but the insight from Mark will do (THANK YOU MARK).
Thank you for your help.
you may like to read → https://support.dxo.com/hc/en-us/articles/6758873797917-Will-my-images-be-sent-to-DxO-servers-when-using-DeepPRIME-technologies-or-AI-Masks
Wolfgang,
Thank you. Hopefully, as DxO is under EU (and France) regulations, the statement is both factual and true. Note that some USA firms using similar AI technologies have not been quite candid about such claims, and in fact do use the material garnered from end user licensees to further “train” the AI/s . Is this AI a true neural network AI?
ALSO, you display as Win 10. MS Windows 10 is now “obsolete” with strong pressure from MS to “upgrade” to Win 11. I have not done so having found no easy path. Do you have any practical advice on how to bypass all of the Win 11 security checks that I do not need and simply upgrade Win 10 to Win 11 in place? (I can get a new harddrive, reset to what Win 11 wants, reformat the new drive, and then copy from the external drive my files – but the MS Win registry would not properly copy/update/integrate from MS Win 10 to MS Win 11 and I would need to re-install all of the non-MS software applications that I use … ).
I believe the training to which you are referring is mostly for software that uses generative AI which creates new content such as when removing and replacing objects, backgrounds and skies. At this point DxO has not implemented any generative AI functionality and all their existing AI functionality is processed exclusively on our machines and not on the cloud.
Mark
Interesting …
During summer I was contacted by DXO to provide personal images+*.dop to better understand how PL customers are editing their images - sounds like training of AI etc …
I don’t believe your personal experience is a good comparison.
First, they apparently asked you to provide those images voluntarily which you could have easily refused. Many programs that use generative AI in the cloud automatically uses images for training purposes without asking.
Secondly, they requested images to understand what users wanted to accomplish so they could design tools to meet user expectations not for ongoing AI “training”.
Finally, DxO currently does not use any generative AI tools nor do they process any of their current AI tools on the cloud.
Mark
2 Likes
However, the AI must be trained. Assuming that this is a neural network AI implemented on a GPU, several points seem to be relevant.
-
Does the AI receive further custom training from the images provided by the user on the machine of the user? (No use of remote distributed processing, “cloud”.)
-
If the AI is implemented in a GPU, then there should be a mechanism to partition the GPU, or to use multiple GPUs as provided by the hardware platform and accessible under the operating system environment. For example, my laptop field machine (upon which much of my workflow is done) has a NVIDIA RTX A3000 12GB Laptop GPU as well as an Intel GPU associated with the Intel CPU.
I don’t really care about export times. I’d rather that the output was accurate and that the interface was responsive when I am working.
I get it but if you have lots of images from a shoot to work through then it becomes more of an issue especially if you have a right turnaround. Besides, it’s very unusual to see things go backwards and export times increase with a new version