Severely disappointed with AI masking in 2023Answered
I have been anxiously waiting for this release, and I have to say that I am disappointed. The AI masking leaves MUCH to be desired. The edges are very fuzzy and undefined on every photo I have tried. Most of these are studio portraits with good background contrast. Even using ON1 sample photos I get the same results. If your subject has hair, forget it.
I guess I will have to continue using Affinity photo or photoshop for any photos I need to mask. It's a real shame, I thought they had something here with all the hype I heard and seen.
Yea, I'm not getting it to work properly either and since AI mask is updated to the new version, I can't go back to the old way either so now I have to mask manually.1
This is SERIOUSLY bad. I am trying to select very clearly defined subjects and the masks are unusable. I find that I spend more time now manually masking than I did before. Sad thing is that Lightroom, photoshop and Affinity Photo all mask these same photos very precisely. I like On1, but this may be a game changer. I have been trying to decide whether to use ON1 (whose interface and workflow I prefer) or go back to Lightroom and photoshop. In Lightroom I can select the subject with one click and 99% of the time it's perfect. With On1 2023, I just tried to select the subject (a basketball player) and all the edges bleed in to the surroundings and it missed the players hands and feet. Funny though, it did a better job at selecting individual spectators in the bleachers who were small and out of focus. Put the same photo in Lightroom and click "select subject" and got a perfect mask.2
I swear I think it gets worse the more I use it. I used Super Select to select a person. The entire person was selected in red, then blue when I clicked on them. When I applied a black and white filter, only their big toe was masked. No kidding, out of an entire person selected, only the big toe was masked.. On another person, the entire person was selected and was highlighted in red, then blue, but only the tattoo on her arm was masked. On1 has a long way to go in my opinion.1
I cannot recommend highly enough that you watch the training and sneak peak videos that have been produced over the last couple of weeks and learn how the new features work. Downloading and reading through the User Guide would be another very good idea. Check out ON1's YouTube channel for the videos.1
This must be a machine and or graphics card issue; Masking is working extremely well on my machine.....I'm beyond impressed.
Dell G3 Gaming Laptop
Processor Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz 2.21 GHz
Installed RAM 16.0 GB (15.9 GB usable)
Nvidia geforce gtx 1050
System type 64-bit operating system, x64-based processor0
I have watched every video made leading up to the release and read the user manual. Apparently I am performing the correct steps, but my results are well below acceptable. The only result so far that I would consider anywhere near acceptable was on a portrait taken against a plain white background. Even then, when I used the same portrait in Lightroom, I got a much crisper, well defined mask. After listening to all the reviewers over the past few weeks, I was expecting much more. But, I find I have to do as much manual masking as I did with version 2022.5. I prefer ON1 over adobe, but the results I get are no where near on par. The masks and the additional cleaning up might be acceptable if no other option was available, but it appears that adobe is far superior when it comes to masking.
FYI, I am using an M1 Mac mini with 16 Gigs of ram. Maybe this is just a Mac version issue. Only time will tell. MY subscription is paid up for the next year, so I'll keep checking back to see if there are any improvements.1
OH, I forgot to add, I am not a newbie to this. I have been taking photos and editing them for many years using both On1 and adobe. I hate Lightroom's catalog feature and prefer to use ON1. Many of my photos involve compositing, which requires masking.0
LOL. I just tried Super Select on a photo that I tried about 30 minutes ago. On the previous attempt, the mask edges were very blotchy and inaccurate. Also her legs(which were selected according to the red and blue high lites) were completely cut off in the mask. I just tried it again with the same photo and same steps and got a very good mask. Edges are ver well defined and it didn't lose her legs. Don't know why. MY steps were exactly the same. I guess you have to expect some glitches in new software. I used to design software but my programs always worked perfectly every time (LOL, just kidding)0
Add me to the list of those having serious issues with the AI masking in PR 2023. For instance, I pick the Add Filter option and select People from the 'Apply with Mask' section. It appears to be masking the person correctly, but when I actually select and apply a filter, large areas within the subject that should have the filter applied, do not. I'm not expecting AI miracles and I'm not picking photos that should be more challenging than what I've seen in the preview videos. I just can't imagine anyone would be remotely impressed with the performance I've seen. So given that there are others who are happy with the masking performance, it must be something related to specific machine configs.
I'm wondering the issue is tied to the Nvidia GPU. issue that has already been identified by On1. I don't have one of the various RTX 'Super' cards (I have a RTX2060), but I have the same problem where I get black previews when I try to use Nonoise AI or TackSharp. And the app has crashed on me half a dozen times today when in the edit module. FYI I'm running Windows 11 on an Intel i7-8700k with 32gb of memory. Never really had any issues with prior releases on Photo Raw.1
Gordon, it's AI, perhaps your download needed to be awakened a bit! LOL1
I have to agree the masking is very hit and miss in a bad way.
1st image is really bad and 2nd is ok but could obviously be better giver the flat background.1
Jim Durfee, I think you may be right. Now, masks seem to be working better. I don't know why, I haven't changed anything I'm doing, but two photos that were masking horribly earlier are masking properly now. Perhaps my computer has issues with change, like me. LOL0
Gordon, I've been running it on an M1 Mini for beta testing so it isn't an issue with the M1 systems.
The videos I've seen have mentioned that the masks created do bleed over and do not create hard edges. Scott Davenport says in one of his videos that this is an undesired condition as it leads to highlights and other masking issues and image artifacts.
This is 1st generation for this technology so cut some slack. Send the company your raw files where you think it can be doing a better job so they can give the AI engine more training and make it better. I've had to send them some that I got poor results from but picking a blue aircraft out of a blue sky can be tricky where they have the same colors and tones and there are no hard, contrasting edges to be easily identified.1
I don't know why, but the masking is suddenly working much better now. I did notice a lot of network activity after I performed the install, but I didn't receive any update notice so I don't know if anything was being downloaded or not. All I know is that immediately after the install the masking was horrible and now the masking seems to be very accurate. Even hair is masking much better now. This is the upgrade I was expecting. I don't know what was going on earlier, but it seems to have resolved itself.1
That's great Gordon. Thanks for the update.0
From a response I got from On1 during beta testing I believe that the AI masking works in two phases. There is an initial analysis of the image that roughly breaks it down into regions, then when a region is selected and mask created from that the initial version is refined to produce the actual mask.
The red/blue previews are based on this initial analysis and may not be a good representation of the actual mask that will be produced by refining. In one of my images people were recognised and given their blue preview but then totally refined out of the final mask. I reported this and sent the image to On1 and their response was that refining needs more work.
The use of a rough initial preview makes sense since the analysing of an image is done when it is first loaded, so producing a really accurate mask at that point would add to loading time and be wasted if masking is not used, but the pre-analysis speeds up use of AI masking if it is used. Something of a balance between speed and initial accuracy.0
If the edges of the mask need to be sharper then you may find that opening up the full mask pane on the filter/adjustment by clicking on the mask thumbnail and experimenting with the Levels setting may sharpen the edges. Try moving the right hand and centre buttons to left.0
I also am finding the AI mask to be far from ideal, almost all the time. For example (admittedly a difficult subject, but I think it could be handled much better), with this original image
Here is the mask as presented in Super Select
And here is the final mask.0
Yup, it seems to work well with some photos now, but not on others. Portrait are masking pretty well now as long as the background is plain and there is a lot of contrast between subject and background. I have a photo though that I took at the marina in Kenosha Wisconsin that it absolutely will not mask. The photo is well defined and sharp. There is clear contrast between the water and sky. When I select the water, the red a blue highlites select "most" of the water, but then when applied only a very small portion of the water is masked. When I try to select people in the image the masking is absolutely horrible and looks like something a 2 year old finger painted.
I have to admit though, that the portraits I masked properly now, were masking in the horribly finger painted style when I installed 2023. So, I am hoping that the masking gets better on other photos as well.1
BTW, I never expect any mask to be perfect. All masks need some refining. But, the selected mask should at least be close. I always expect to have to refine hair, etc, but when selecting a large body of water or the sky it should be relatively close. I have seen improvements in some photos since installation yesterday so maybe time will improve.0
I have to admit that this release is a marked improvement over 2022. I haven't had to "force quit" the program at all, there are improvements to sky swap, and the masking, even though it still needs some improvement, is a big upgrade over 2022. I have been waiting for this release to make a decision to stick with ON1, or with Lightroom. I have been using both for years and I need to get rid of one. I am going with ON1 mainly for the non restrictive cataloging. With Lightroom's catalog, I have to keep my catalog and files on an external drive and swap it between computers as I work. In a studio setting, swapping from tethered computer to my editing computer to my laptop was a big hassle. With ON1, the files are saved directly to a NAS drive which can be accessed by all the computers.
I am going to stick with ON1 and I think we will continue to see improvements with it over the next few weeks.0
Super Select AI is a great step ... backwards!
It is a great idea, but most of the results are really terrible. It cannot even create a good mask for some objects with a green photo background.
With the AI Quick Mask - Tool from version 2022 this could be done within some seconds. But this great tool is missing now.
So we have to wait some months for version 2023.1. Great promises, but a bad result. I am disappointed.0
Here is an example of what I would call a success. As you can see, this photo is out of focus and has a very busy background. Super Select selected my granddaughter pretty well and also selected part of the lawn mower. This did not surprise me though. With a little bit of clean up, I got an acceptable mask.
It does seem to be hit or miss though. Some photos that I consider clean and easy to mask are not masking well.0
I've gone back to Photo Raw 2022 for now. I'm finding AI Super Select unusable, and I much prefer the old version of NoNoise AI; the new version is over the top for my taste. I use both Windows and Mac.0
Here is a sample of using an adaptive preset to add a brick background to the photo. The original background was a plain gray wall. I haven't done any refining to the mask. So, on some photos like this it works really well.0
Finally got around to screenshooting a failure. In this photo, the selections are pretty good, but the masks are unusable. This same photo is easily masked in Lightroom, photoshop and Affinity Photo.0
Just playing AI mask since monday release; quite disappointed, here is an example, select architecture (as well in tools than add a filter) in a clear raw photo, pilars are not entirely detected even if color, contrast, geometry are same as walls etc.
Can't find a way to add the pilars lacks in the mask, asking for "add" just add sky, or trees etc, it would be much more interesting to have the control of expand (paint again?) the mask "architecture" (or others AI mask). For people, even after many tries, mask is too big, I tested again and again with a photo where the person is the main subject, it appears like David Kick bird. Very difficult too understanding the workflow needed for get layer(s) generated and how to have control of human brain in this process...0
You can't paint in (from within AI Mask) to add missing parts. You are limited to just adding or removing areas identified by the AI.
I think it is a BIG mistake to completely remove the ability to paint in and paint out as you could before with Quick Mask.
I would encourage everyone who is disappointed with the removal of the option to paint in and paint out to complain to ON1. This worked once before when they (as I recall it) removed the clone stamp when they added the healing brush. I would use the in-program option Help - Provide Feedback
Until then, you can let the AI masking do its thing and then refine the mask afterwards. In cases like yours, where the AI did most of the work pretty well, it would be a minor touch-up to the mask afterwards.0
This was the reply to my request to get AI mask back, so the more that write in, the better chance we have.1
I sent in the following
I find the new Quick Mask in 2023 to be (far too often) very imprecise. While I can see its utility in some images, I think it would be good to ADD back in the ability to use the old method of Quick Mask, where you get to choose areas to add or drop.
Maybe it could be under the "Mode" drop down;
- AI in
- AI out
- Paint in
- Paint out
And I got the reply;
"Thanks for your feedback. I'll pass along your comments to our product team for consideration."2
Please sign in to leave a comment.