Leveraging latest AI capabilities for future developments

DxO PhotoLab Feature requests
1

I see very interesting things coming when it comes to AI, in the future of my practice.

First of all, I don’t talk about the generative AI bunch of instragram-like nonsense that is trained on unlicensed imagery databases…this there is litteraly a shitstorm of IP rights infringement lawsuits coming in to discipline this mess, and litteraly everyone will have to retrain their model from scratch and paying for what they train on (well except social media, because you are willingly submitting your pictures to them).
Sidenote : look at the very small lines of Adobe Terms and conditions regarding the real IP ownership of your picture if you include content generated with Adobe Sensei engine…it may surprise you.

I am talking about what AI and deeplearning model training could do for us : I have several suggestion to submit, your feedbacks are welcome :

  1. face recognition for local adjustment : I can spend literally hours doing the very same local adjustments on a face…in 100 different photos (for the bride in a wedding for example), it takes forever, and when I remove a small pimple or a blemish on the face, I have better not to forget it in the 47th photo that comes along the way. Let’s call this “AI adaptative mask”, it could adapt the zone of the retouch (if the bride is looking in a different direction), the intensity if needed, depending on other parameters.

  2. There is a tool coming in for Lightroom (from a third party) that will look at all your database, identity your “style” of retouching, luminance, color temp, cropping aspect, and will propose something when develop a raw. It should be adaptative, and you will have an incentive to train your own model by correcting the proposal, categorizing it along the way (landscape, portrait, portrait of a given person which can be different because you always process the portrait of your wife differently than of your distant relative) . This is the best incentive by the way to keep customers loyal in time. Databases should be in local of course. The model would train on the huge database that we all have, and on the thousands of dop sidecar files that we all have. The data is there, up to DxO to make something useful out of it.

The idea is the same, productivity, and having AI or trained deeplearning engines proposing things that could save us 80% of the time we spent repetitively doing more or less the same thing. And since each image is unique, we tune it this way, but a software could definitely identify quickly the retouch that we always do in each image, the “look” that is our signature, and adapt the unique context of the image (luminance, face in a different postion, color temp, saturation…) to reach that “look”. Effective blend of machine and human contribution if you ask me.

2

There are already a number of proposals with regard to AI functions and development there will continue to progress steadily. At the moment, as far as I know, it is not possible to predict at what point the investment in AI computing power and the price that customers are prepared to pay for it will intersect. Most models are currently moving towards the purchase of online computing time.

Black Forest Labs has taken a first step in the direction of your requirements with a generative AI that allows you to upload your own images and generate new images on this basis. If I transfer this to Photolab, the Photolab AI functions should be able to distinguish between different customers and several corporate identities - reproducibly. So from a helicopter perspective, you could certainly find a few more requirements.

The existing AI systems have, as you pointed out, used image sources that are legally disputed. What image data should DXO use? For a basic model, you will obviously need a lot more image data than what is in your database. Are you prepared to make your database available to DXO in order to create a sophisticated basic AI model together with other users? Not only you but also other users would benefit from this. Do you want to be compensated for this or is it a cooperative model? Many things are still unclear and therefore uncertain.

1 Like
3

It’s a good point.
The data base is not the pictures themselves, nor is the sidecar files, and there is definitely value in the data that we longtime users own. I don’t want to dispute the exact property status of the .dop files generated by my usage on my data, but from a software that is not my property and for which I only own a licence of personal use.
However the pictures are mine and in no way I am granting any right to use them to train models on my work, that’s why I don’t use any social media like instragram or facebook.

A clever thing would be to “monetize” the willingness of the users to, while using DPL, would be ready that the results of image analysis and how we process the image, could be shared and sent to feed the training of a global model. Maybe a good way to earn a few free licences during the beta testing campaigns.

No one has the perfect answer currently, spending tons of money and electricity ressources to train a model, or distribute the effort to the large client base. Probably a bit of both, and DxO has the opportunity to set a first landmark.