First, second and third impressions of Luminar 4

[purplehaze]

First: Wow!
Second: Hmmm...
Third: A useful plugin for Capture One, but not a replacement.

Pros: The AI features are very impressive, especially the intelligent masking of skies and the skin smoothing and face light tools. The sky masking is a huge time saver that is hard to argue against.

Cons: No DAM tools. Maddeningly slow, with multiple crashes on my 4 GHz Intel Core i7 processor with 24 GB 1867 MHz DDR3 and AMD Radeon R9 M395X 4096 MB graphics card. No colour proofing. Let me say that again: NO COLOUR PROOFING. Right there, that relegates it to the plugin category for me Also, I had no colour consistency between file formats until I reverted to my display's native calibration. Now it seems I cannot retrieve my Spyder calibration through System Preferences. What gives? Many (most?) of the filters or tools have that impenetrable black box quality to them, which can be frustrating. Manual masking is much less refined than in Capture One. As I mentioned in another post, use as a plugin with Capture One is more cumbersome than it is with Photoshop or Lightroom.

It has been fun playing around, but now I need to get up to date on Capture One.

Edit: I figured out how to get my calibrated profile back, but am still wondering why it seems not to work with Luminar.

[Manfred M]

Janis - one has to remember the purpose of the tool; it is a plugin to enhance the normal editing tools one uses. It like many other plugins (the DxO Labs Nik collection, Topaz Studio, etc) it supplements rather than replaces your standard editing software. It's one of the few plugins I've bought and use.

[purplehaze]

Janis - one has to remember the purpose of the tool; it is a plugin to enhance the normal editing tools one uses. It like many other plugins (the DxO Labs Nik collection, Topaz Studio, etc) it supplements rather than replaces your standard editing software. It's one of the few plugins I've bought and use.

Agreed that it works best as a supplement, Manfred, but it is being marketed as a standalone editor, as well as a plug-in. I don’t think it hurts to warn people just starting out in processing that it is not the be-all and end-all some advertising and influencers would lead one to believe.

[DanK]

Janis - one has to remember the purpose of the tool; it is a plugin to enhance the normal editing tools one uses. It like many other plugins (the DxO Labs Nik collection, Topaz Studio, etc) it supplements rather than replaces your standard editing software. It's one of the few plugins I've bought and use.

What do you use it for?


Sent from my iPad using Tapatalk

[Manfred M]

What do you use it for?

Sky replacement; quick with good results both using the internally generated skies and ones in my collection of sky shots. For example:

The Three Muses of Theatre


I show the before and after image. Truth be told, it would have been an easy shot to do by hand too.

[DanK]

Sky replacement; quick with good results both using the internally generated skies and ones in my collection of sky shots. For example:

The Three Muses of Theatre


I show the before and after image. Truth be told, it would have been an easy shot to do by hand too.

Thanks. I asked because apart from filters, by which I mean canned effects, there is a great deal of overlap between what various software packages can do. I have virtually no interest in filters, so my interest in additional software packages is primarily whether they can do something quite different. I don't often replace skies, but this would be an example. Another example is the control point technology in Nik, which is fundamentally different from any other selection tool available in the software I have.

BTW, I have no expertise in AI, but I really wonder what vendors mean when they say their software uses AI. I have a hunch what they really mean is algorithms based on deep learning, not AI. AI, as I understand it, means that the software continues to learn. E.g., it would learn over time what YOU are trying to do when making adjustments. Deep learning, again as I understand it, is more or less canned AI. For example, Canon refers to the AF system in the new 1D as being based on deep learning. That is, they trained the software using a huge library of images--essentially, a form of AI--but what is ported to the camera is the RESULTS of that AI. The camera doesn't continue to learn. I have an ill-informed hunch that when editing software vendors reference AI, they really mean deep learning and don't mean that their software will continue to learn after you install it.

Speaking of such things, I just tried tried a demo of Topaz Sharpen AI on a photo that was a bit soft because of a user error. I was impressed. it is somewhat slow--every time I moved the frame to look at a different area, it had to recalculate--but on my computer, it was seconds, not minutes. The results weren't as good as a truly sharp photo, but they were considerably better than I had achieved with any other software.


Sent from my iPad using Tapatalk

[purplehaze]

Dan, your hunch is correct that this software does not actively learn from you. However, I would make a slight correction to your terminology. Something that is based on deep learning may be “canned”, but deep learning itself is adaptive. I know something about this as it is currently having a very big impact on my profession, which is translation. The very best machine translator available right now learns from its users and it is in fact called DeepL.

From Techopedia:

Deep learning is a specific approach used for building and training neural networks, which are considered highly promising decision-making nodes. An algorithm is considered to be deep if the input data is passed through a series of nonlinearities or nonlinear transformations before it becomes output. In contrast, most modern machine learning algorithms are considered "shallow" because the input can only go only a few levels of subroutine calling.

Deep learning removes the manual identification of features in data and, instead, relies on whatever training process it has in order to discover the useful patterns in the input examples. This makes training the neural network easier and faster, and it can yield a better result that advances the field of artificial intelligence.

[DanK]

Dan, your hunch is correct that this software does not actively learn from you. However, I would make a slight correction to your terminology. Something that is based on deep learning may be “canned”, but deep learning itself is adaptive. I know something about this as it is currently having a very big impact on my profession, which is translation. The very best machine translator available right now learns from its users and it is in fact called DeepL.

From Techopedia:

Janis,

Thanks. Sorry I was unclear. My point was just the timing. When I read "AI", I take it to mean that the machine's adaptation is continuing, and that isn't what these software packages offer.

Dan

[purplehaze]

Many people understand it as you do, and accuse companies like Luminar of misrepresenting their product, but the term "artificial intelligence" has multiple meanings, and, in any case, the concept of "intelligence" appears to be fluid:

(From Wikipedia)

As machines become increasingly capable, tasks considered to require "intelligence" are often removed from the definition of AI, a phenomenon known as the AI effect.[3] A quip in Tesler's Theorem says "AI is whatever hasn't been done yet."[4] For instance, optical character recognition is frequently excluded from things considered to be AI, having become a routine technology.[5]

Sorry, but I'm a word nerd.

[Manfred M]

Dan - your assumptions are correct, so far as I understand the technology.

In machine learning of this kind, two things are introduced to the algorithm; the state of the input image and an image showing the desired outcome. This generally means throwing a lot of images at the algorithm; measured in the thousands to hundreds of thousands. The learning algorithms figure out how to get from the input state to the desired output state. That way if a random image is introduced, the machine learning algorithms are applied to it to create the output state. Often the results are quite good, depending on how closely one looks. The limitation has always been the number of artificial neurons available, so some of the graphics cards manufacturers, specifically nVidia, have thrown a lot of resources into that field. These cards perform a lot of very simple calculations to create high frame rates for video games and have proven to be very effective in the machine learning field. We just get the benefit of the output from these algorithms.

The software provider can throw additional data sets at the algorithm in an attempt to improve the results. In machine learning, the general rule of thumb is that the more data was fed into the algorithm the higher quality the output tends to be. The software we use (Topaz AI, for instance) does not have the machine learning interface. We merely benefit from the data sets Topaz has thrown at it. Since I bought the software (probably around 8 months ago), Topaz has updated it several times.

Janis - you might be interested to know that in Machine Translation, some of the data sets have a very Canadian base. As Canada is an officially bilingual country, the Federal Government works in both official languages. Hansard (essentially the minutes of the daily sessions of the House of Commons) is a very large and well translated (English / Freench) data set that a lot of machine learning groups have used for input to train their algorithms.