Since feedback was requested, here are my considerations.
The idea is potentially good, but unfortunately, many complexities are evident both in the final results and in the application. However, these are inevitable when developing new solutions.
The results are reasonably decent for standard images taken in typical shooting conditions. If we stick to the scenes used for the presentation, there are interesting corrections on macro photographs of flowers or insects, and on well-balanced images. If considered as a product aimed at an inexperienced and amateur audience without particular shooting needs, it may be acceptable, although the reference models for the images are extremely limited, and the analysis of the obtained results does not seem accurate. If these shortcomings are overcome, it could lead to acceptable and already interesting results.
The process tends to flatten the images, stripping them of their structure and depth, which means it is unusable for high-key or low-key images, in product still life, for portraits taken in a studio or on location, in landscapes with strong visual impact, and in shooting situations where depth and volumes are essential. The approach is quite simplistic in this regard.
Unfortunately, nowadays one would expect sophisticated depth and structure mapping of the image in order to generate corrections based on a three-dimensional logic. But as can be seen from how the Lens Blur panel works, Lightroom still lacks the ability to map the scene for volumes and subjects. This issue affects the correction of high-light and low-light areas: a shadow caused by a decrease in light is perceived as a dark area and is balanced accordingly, rather than based on the volume of the scene. These limitations are still evident in all AI applications.
Therefore, if the goal is to create images with a high dynamic range, we are really not there yet. When working in color grading for professional HDR solutions, readability in shadows or highlights is not considered the final element; instead, the aim is to extend the dynamic range as much as possible to deepen the scene. Thus, in this case, the application of corrections is simply terrible. The images become flat, structure is lost, and above all, the photographic intent is lost.
Errors are obvious on high-key landscape images such as snowy northern landscapes, studio photographs with deep shadows, and portrait photographs with pastel tones. The aesthetic reference standards are too limited, and the application often produces worse results than the original image. Verifying files from different cameras, photographers, and manufacturers, I can state that, on a very limited sample of about 300 images, only in 10% of cases was the result potentially interesting.
The main limitation is in the application to the Adobe Standard color profile alone, which, by its very design, does not allow the full gamut of the sensor to be rendered, with obvious difficulties in reds and blues. It is also impossible to select a 3D LUT within the profile to properly correct the starting base. Other main limitation of this solution, besides the extremely limited data source, is the inability for the photographer to allow the app to train on their own files, thus providing the ideal working base for making corrections.
If the goal is solely to normalize the image in order to start working on it, this process often makes the photographer’s work more laborious, as in most cases the starting point for their work is already the image they have created.
I hope to soon see a more accurate development of the product and be able to apply it proactively to a large number of images. I trust that there is the ability to create a new tool that enables professionals to integrate a series of solutions that allow for maximum customization of the individual photographer’s experience. More than a multitude of generic references, it would be useful to have a private and non-shareable AI training area dedicated to the individual photographer, as well as the option to use a specific color profile and proprietary 3D LUT that can provide the correct mood to the image. Obviously, in color grading, the AI should be trained to maintain the deviations expressed by the LUT so that specific tones of the image are not corrected. This approach would better assist all photographers and allow reportage photographers to maintain compliance with the C2PA standard.
Note: to provide accurate feedback, I tried modifying the starting data of the adaptive XMP profile by inserting different references for color profile and look-up tables, obtaining better results compared to the initial setup, but with clear limitations imposed by the AI control system.
49
Replies
AdChoices