Can and will the Ai features get better?

[LostInTranslation]

Hi,

i have not upgraded to 2.6 yet, I always wait until my current projects are finished in case something is not working.
But i have watched a couple of videos. It seems that the new Subject/Object Selection Tool is not getting great results most of the time.

Affinity stretches, that all processing is done locally and none of my artwork is used for training.

Does this also mean that we are stuck with what it is now?
How could it get better? Would Affinity need to train another, better model?

[carl123]

1 hour ago, LostInTranslation said:

Does this also mean that we are stuck with what it is now?
How could it get better? Would Affinity need to train another, better model?

From my understanding...

Affinity uses 3rd party pre-trained models for these 2 new functions. If the 3rd party improves those pre-trained models, Affinity may include them in new updates just like it does for camera and lens data

Affinity's implementation of these pre-trained models into its apps may improve over time leading to better/faster performance

For Windows users only the CPU is currently used with these models. Serif are aware and are working to get the GPU supported as well. If/once that happens there should be a performance boost for those of us using Windows. Not sure is using the GPU also improves the quality of the results or not.

These ML Tools appear resource intensive so users may wish to consider upgrading their PC's components (e.g. ram, video cards etc) should they have need of faster performance or investing in a new computer if their needs/costs justify it.

 

1 hour ago, LostInTranslation said:

It seems that the new Subject/Object Selection Tool is not getting great results most of the time

  

I doubt it would have been implemented if that was true. The results do vary and may require additional tweaking but for some users it's far easier/faster than the previous alternatives of using the Selection Brush Tool or the Pen Tool etc. Bear in mind that even the industry giants like Photoshop also fail to make perfect selections all of the time so the technology still has room for improvement. For now, it's just another tool we can choose to use or not in our respective workflows

 

 

 

 

 

[Granddaddy]

You can get a good idea about the inadequacy of AI selection and masking tools from Robin Whalley's comparison of Photoshop and Affinity Photo 2.6 published this week at

https://www.youtube.com/watch?v=9xipGsC8j7I

He concludes that APhoto is slightly better than Photoshop at AI selecting. He also concludes that neither is very good.

My experience with APhoto 2.6 and On1 Photo Raw 2025 is that AI masking and selecting is adequate and even fun for quick and dirty work when touching up and enhancing family snapshots and hobby photos. AI sloppiness isn't much of a problem for such routine work.

For more precise work with fine details, AI tools are grossly inadequate. Fine work requires you to use traditional selection brushes and pen tools. This is easily seen if you use AI to remove a complex background around people. Once the subject is isolated, insert a white background to see all the spillage into the area that at first appears transparent. There remains a lot of manual cleanup to be done following application of AI selection tools.

Whalley refers viewers to one of his earlier tutorials on making fine selections using traditional tools. They work far better than the AI currently available in any product. The traditional tools actually take very little time while getting far superior results.
https://www.youtube.com/watch?v=PUmwGYcWdk8

I'm less and less interested in AI hype. Vendor demonstrations and even many end-user tutorials invariably use carefully selected photos that are unrealistic and even irrelevant in my world. They also tend to ignore sloppiness that is apparent even to my aging amateur eyes.

I can add that another AI tool (generative fill in On1 Photo Raw) has never improved any photo I've tried it on despite assurances by the demonstration videos from On1. I've had better results with APhoto's Inpainting than I've gotten from the OPR Generative Fill.

 

[PaulEC]

My experience so far with these ML tools is that they are OK for simple selections, which would be quick to do manually anyway, but no better for more complex selections. I had hoped that they would be useful for things like hair, multiple small cut-outs, such as leaves and tree branches and semi-transparent objects. Unfortunately tidying up selections after using the ML tools takes just as long, if not longer, than making the selections with other selection tools, and the finale results are no better.  

[Granddaddy]

Take Better Photos has a detailed analysis of the pros and cons of AI selection tools in Affinity Photo 2.6 in comparison to Photoshop. The latter has a one-click solution for fine details like hair.  APhoto's AI selection requires additional work with the refine brush in such cases, but it is still quite useful. Best of all, the APhoto AI selection tools are a free update for those who purchased APhoto 2 at any time since it was released in November 2022.

https://www.youtube.com/watch?v=Gmid3iDTVmI

 

[Pšenda]

17 hours ago, Granddaddy said:

Best of all, the APhoto AI selection tools are a free update for those who purchased APhoto 2 at any time since it was released in November 2022.

Unless they happen to have macOS with Intel, whose users feel ripped off and robbed 🙂

[R C-R]

5 hours ago, Pšenda said:

Unless they happen to have macOS with Intel, whose users feel ripped off and robbed 🙂

FWIW, I have an Intel Mac & I do not feel ripped off. Of course, I would like to have the 2 ML features but since just about everything else works fine, I can live without them.

[LostInTranslation]

Thanks to everybody for their insights.

Right now the AI-feature seems to be a better (or at least faster) selection brush, but still needs work most of the time.
I have no problem with that, but on the other hand it is not a "must have" for me.

I wanted to try it on my iPad, but apparently I "do not have Apple silicon" so it does not work.
I will update on PC soon and then hopefully get to try it myself.

[Medical Officer Bones]

For fun I tested Photoshop, Affinity, and Krita (with the free open source segmentation selection plugin).

Photoshop 2025 wouldn't cooperate at all. Soft edge, hard edge, ... Unusable.

In Affinity Photo I had to turn off soft edges. Then it did an okay job.

Krita was the simplest. No change of settings required, a single click on the wood sufficed. The selection is slightly better than Affinity's.

Neither Affinity's nor Krita's selection is acceptable 'as-is', and both require more (manual) work.

But I can see the potential to at least speed up the selection process.

[Pšenda]

11 hours ago, R C-R said:

FWIW, I have an Intel Mac & I do not feel ripped off. Of course, I would like to have the 2 ML features but since just about everything else works fine, I can live without them.

For me, the only interesting thing about the current ML functionality is that it's the first time that external modules (plugins) are used. This means that Serif has already made progress in implementing them, and that they will hopefully be released for wider use soon. I would guess that Blend Tool could use them as well.