More Generative AI - Sadhu + Tigers, etc

[Manfred M]

I've been working with a different generative AI model and created this Sadhu on the ghats in about 45 minutes. This one is called MidJourney and is a pain to set up, but it can do strong photo-realistic images if coaxed properly, some of the time.

I am planning to add some more of my playing around and musings to this thread, rather than creating a lot of new threads, jus to keep things togehter.

I'm starting to get the hang of it but need a lot more practice and refinement. Some work in post on this image.

The main issue with this software is that the maximum resolution is 1024 x1024 (for rectangular format), so I upsampled with Topaz Gigapixel AI.

[Manfred M]

[Cantab]

Manfred, at the risk of asking a perhaps stupid question, I'm curious about which aspects of these two photos were AI generated. In the second photo, Midjourney possibly provided you with two tigers?

[Manfred M]

Manfred, at the risk of asking a perhaps stupid question, I'm curious about which aspects of these two photos were AI generated. In the second photo, Midjourney possibly provided you with two tigers?

100% of these images were generated by the software. I did a bit of dodging and burning to improve the blending of the elements. Nothing was added, but I did remove a bit of bright white garbage from near the Sadhu's foot.

The tigers were a bit too "crunchy" so I softened the fur.

[MrB]

Looking at the fur on the body of the left tiger, surely the head of the right tiger should be out of focus?

Philip

[Manfred M]

Looking at the fur on the body of the left tiger, surely the head of the right tiger should be out of focus?

Philip

I agree that there are a number of noticeable "defects", especially in the tiger image. The fur is too distinctive and a bit too "crunchy", the stripe patterns of both animals is almost identical, the DoF on the two animals is not quite looking right and the light on the right-most animal does not quite line up with what we see on the left tiger.

The Sadhu's left foot's heel is a little strange, but then a number of the Sadhus that I saw in real life had some physical deformities.

Overall, photo-realistic images are still not quite there yet with this technology, but if one is looking for graphic art, even to the level found in graphic novels with their higher level of stylization, the quality there is better. Here is an image generated when I asked MidJourney to show me what the Midjourny Bot looks like:

[stuck]

Interesting that the face is obviously female.

[Cantab]

100% of these images were generated by the software. I did a bit of dodging and burning to improve the blending of the elements. Nothing was added, but I did remove a bit of bright white garbage from near the Sadhu's foot.

The tigers were a bit too "crunchy" so I softened the fur.

Manfred, how much input did you provide for the two images. "A sandu on stairs"? "Two tigers"? Or were you more specific?

What fascinates me most is the extent to which MidJourney is, or is not, able to create an "artistic" image (aside from technical photographic imperfections). Or put differently, the difference between a bot produced piece of writing and the same topic/issue dealt with by a truly gifted human writer.

[Manfred M]

Interesting that the face is obviously female.

Each prompt generates four different images. I just happen to have like that one best. Here is one of the others that is quite different from the first.

[Manfred M]

Manfred, how much input did you provide for the two images. "A sandu on stairs"? "Two tigers"? Or were you more specific?

What fascinates me most is the extent to which MidJourney is, or is not, able to create an "artistic" image (aside from technical photographic imperfections). Or put differently, the difference between a bot produced piece of writing and the same topic/issue dealt with by a truly gifted human writer.

The prompt that gave me the Sadhu (after a number of different attempts) was "photo of a sadhu sitting on steps of a ghat with steps filling the background Kodachrome 200 film". An image with a growling tiger was never generated and the closest it came to that was having a tiger with its mouth slightly open in another image.

The tigers used: "photo of a growling tiger, with her cub, standing on the bank of a river looking through the grass and looking at the camera using 600mm lens at f/4"

It took me several tries to refine the text to give me what I was looking for. Some of the results were quite far off what I had planned for the image to look like.

In one case, I simply entered the word "Theyyam" and I got a very usable image in one try. I tried several other refinements and ended up working the first output in Photoshop.




I tried to see how close I could get to my recent postings of white pencil crayons in a vase on a white sweep:

White Vase + White Pencils






I gave up because MidJourney was not even coming close to what I wanted. The lack of control and relatively strange output suggest that very specific work is still rather hit and miss. On top of that, photorealism is not always possible.

It's fun but in my view, the technology is still not where it needs to be for strong creative work.

[Manfred M]

I tried using Midjourney to create a landscape (really more of an urban landscape) of what I saw last winter while visiting the famous houseboats on Dal Lake, in Srinigar, Kashmir.

This is more or less the best I could get our of Midjourney - it looks more painterly than photo-realisitic.






This is the image I was trying to get close to.







I'm not really happy with the Midjourney result, although it is a nice image. It does show some of the current limitations of the software.

[Cantab]

Manfred, does Midjourney accept instructions about the type of lighting, etc. you want? Or does it simply focus on the subject matter?

[Manfred M]

Manfred, does Midjourney accept instructions about the type of lighting, etc. you want? Or does it simply focus on the subject matter?

It does. In this scene I asked for winter and overcast skies. In others I have made suggestions on light direction. The problem with this technology is that it parses the text and one can never be sure of which aspects of the data it will use to generate the image.

As an example; flash, fill light, softbox, studio lighting did not make any difference to what it generated.

The text that went into Midjourney to generate this recently created scene was: "photo in winter of houseboats on Dal Lake in Kashmir with mountains in the background sky, is heavy overcast gray "

[pschlute]

Quite fascinating, and an area of image creation that will continue to take-off massively in my view.

It is not photography, but it certainly is art.

[rpcrowe]

The AI that I have seen also looks a bit painterly. This video was on YouTube and states that it asked AI to create a portrait of a beautiful woman for every country.
https://www.youtube.com/watch?v=1qGT9WjNN0k&t=205s

I enjoy these but if they were photographs, I would offer that they are a bit over processed - eliminating skin texture.

[Cantab]

Richard, I had a look at the video. As you noted, the AI algorithms resulted in the elimination of all skin texture. There was something else about the images that I can't quite put my finger on. There was a similarity of style throughout many, perhaps most of the images – in addition to the skin texture issue. I suppose it may be the result of whatever artistic styles or values had been programmed into/taught to Midjourney.

[Manfred M]

Midjourney seems to, by default, be creating output that has more of a painterly than photography look and that is certainly true in the latest release or two. That "graphic novel" look seems to be quite popular with a lot of users.

There is an instruction that can be added into the text sting fed into the "engine" that will reduce this stylization. Adding the string --style raw gives a more photo-realistic look, which is definitely present in my Sadhu image. Unfortunately, it does not always work as I tried that with the houseboat scenes and it still looks painterly.

I suspect that the creator of the video wanted that look, because it can be downplayed.


I created this portrait with heavy bokeh with the raw switch with Midjourney 5.2 and the image is more photo-realistic.

[llewin2014]

Good day, Manfred. What are these types of documentary-like imagery used for? How would this type of image be categorized, and more importantly, how would the work be presented in exhibition or online, or within, perhaps research?

Thank you.

Lance A. Lewin
Photographic Society of America (PSA) Global B&W Photography Director
International Visual Sociology Association (IVSA)
IVSA Code of Ethics North American-Canada Think Tank

[Manfred M]

Quite frankly Lance, I was just playing around with MidJourney to gain a better understanding of this evolving technology and nothing more. It reminds me of the early digital days when film photographers suggested digital images were NOT photographs...

Let me be 100% clear; these are NOT photographs and should never be represented as such. However, they are images and as such we need to understand that as this technology evolves, they will make their way into more general use. If you look at the Hollywood writers and actors strikes; AI and its use was / is a key issue on the negotiating table. I can for instance, see PSA (or FIAP or CAPA (Canadian Association for Photographic Art, the Canadian FIAP affiliate), which I am a member of) opening a Generative AI division at some point.

I am a CAPA Certified Judge as well as the Competitions Director at one of the photo clubs I belong to, so this topic is of interest in terms of screening and judging images. The policies on AI in images is fluid and evolving. I suspect PSA's views are similar to CAPAs and Generative AI content in images is not allowed in competitions. That seems to be the "easy" part, but other issues are more challenging as a lot of photo editing software has incorporated Machine Learning functions into the software. Photoshop's Spot Healing tool appears to use fairly rudimentary AI, but Content Aware Fill seems to be very heavily based on this technology and the borderline between what should be allowed and what should not be allowed in photo competitions is not clear. Adobe's Firefly appears to step over the line.

Once we get out of photo competitions, things are a lot less messy. Computer generated content and technology driven content are going to have a significant impact in the advertising and social media fields. Deep fakes should be a concern to everyone and there are entities out there, with agendas, that will be exploiting this technology for their own gains. Legal opinions are not going to have a lot of impact on these actors...

In terms of "Art", I will be exhibiting a piece that is about 70% photograph and 30% generative AI at a local gallery later this month. It will be labelled as Photograph + Generative AI. I have no issues using ChatGPT to generate ideas for image titles or Artist's Statements, but I will in all likelihood be revising and editing thesecontent to make it "my own".

The image in #17 in this thread was used as a placeholder for a seminar I am giving in a couple of weeks. I needed an image that clearly shows bokeh, and none of the ones in my library were good enough. Right now MidJourney only generates images that are 1050 x 1050 pixels, so that limits their use to screen based work or very small prints. I have played with some of the upsampling tools (Topaz GigaPixel AI) to see how far these can be pushed.

I did give a short talk on Generative AI at a local photo club yesterday...

If you would like to discuss this further, please send my a PM and we can continue this "conversation".