Taking the AI plunge

MartinLogan Audio Owners Forum

Help Support MartinLogan Audio Owners Forum:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.
When we had this house built I just had them run cat 5e to many rooms. Never know when you might need it. That was in 08.
 
The new wore off quickly for me.
Face swapping was really cool for about 1-2 days. I use MidJourney if I have an idea, but after about 150 images, I felt like I had the gist of it when I want something to play with. I see plenty of example prompts and very interesting artistic ideas in a couple forums I belong to.

CoPilot in Visual Studio is the gift the keeps on giving. It is continuing to impress me.
I'm still learning what I can do with ChatGPT4 in terms of having it find things for me and do things for me. Since they released plugins, there is now a new wild west. I'm finding a lot of the plugins have redundant functionality, but at least they order them by popularity which is one source of ratings.

I did install Stable Diffusion on my gaming computer with a 4090 on it and it works, but I'm still not getting the results I get from MidJourney. I don't think I'll bother with the light ChatGPT 3 equivalent that I could also run locally. I don't see the point. This technology is on an exponential trajectory and there is little point in doing this locally.
 
Been playing a little with MidJourney using very simple prompts and letting it fill in the details in the style of a well known architect, brand, or artist. It's interesting to see how well executed these ideas are. The Louis Vuitton VR headset looks like a realistic product shot that could be used for an advertisement.

It is giving very good results from very different inputs, and it seems like it is starting to handle text a bit better.

/imagine treehouse designed by Frank Lloyd Write
/imagine pink VR headset designed by Louis Vuitton
/imagine H.R. Giger style alien wearing a VR headset melding into its head

treehousebyfranklloydwrite.jpg
louis_vuittonHeadset2.jpg
HRGrigerVRHeadset.jpg
 
There is some debate about just how much effect all those camera-related prompt elements actually impacts the output. Remeber, MidJourney and Stable Diffusion (SD) are not cameras, they are language processors that try and 'pull' the prompted image from the dataset (model). To the extent camera terms help, it's only if images in the data set had related text tags associated with them during the training phase.
ChatGPT uses them because the data set it was trained on makes frequent mention of them, so the myth perpetuates.

It looks like people are getting very good results with a very wide variety of prompts.

I saw someone posting an interesting image with lots of camera settings that made little sense to me. The person posting didn't understand photography at all and just took what Bard gave him.

The prompt specified a wide angle shot using a telephoto lens. It specified an ISO value, but not the camera or film used. It specified an Aperture, but didn't specify a focal length. The image looked like a tripod panning motion shot with a fast moving car and blurred background which is a challenging image to capture, but the prompt specified manual focus. The car was also at a 3/4 view, so the focus would have to change very quickly while panning the camera. Basically impossible to do manually.

I thought about what was specified for a while and the image created. Even though there was superfluous data that I'm sure Midjourney ignored, the rest could be made to work given there is no real camera.

A telephoto could get a wide shot if taken from far enough away. With a car coming at you, being far away would be a good thing. The shutter speed of 1/800s seemed very short for panning blur shot, but considering the distance and it being a telephoto lens, it might work. Also given there is no real camera Midjourney could make that shot work.

So more power for Midjourney from taking that prompt and creating a great looking image with it.

---------------------------
Below I created a more photographically accurate prompt that seems like it could work to create the image that the other guy was after. I'm not saying that I could take this shot in real life, but at least it seems possible. I also notice that at f5.6 the rear quarter panel is out of focus which makes sense. At 200mm f8 would probably be better to get the whole car in focus and because of the panning action the car would still have plenty of separation from the background.

/imagine Ford GT40, panning motion shot taken from a distance with a full car in focus, 3/4 view, high end digital camera, 200mm lens, f5.6, 1/400s --ar 16:9

GT40_panningshot.jpg


I tried the exact same shot changing the prompt to f8 and the rear quarter panel is now in focus.
So I think that the camera settings do actually work and impact your images, but that Midjourney will do its best to make the most out of camera gibberish as well :)

GT40_panningshot2.jpg


One more thing bothered me. The wheel's hubs didn't seem blurred enough.
So I specified the car's speed and dropped the shutter speed and one of the 4 images it gave me showed the hubs blurred. Of course in this image the car is no longer presented in a 3/4 view. So still not exactly what I asked for.

/imagine Ford GT40, traveling 100mph, panning motion shot taken from a distance with a full car in focus, 3/4 view, high end digital camera, 200mm lens, f8, 1/50s --ar 16:9
GT40_100mph_50s.jpg


It seems that Midjourney ignores ISO completely. I tried to get some film grain by specifying Kodak Gold film and saw none, however I did get the warm hues that you would expect with Kodak Gold on all 4 images. I also tried Fuji film and it had the greenish case as expected in some, but not all images.

/imagine Ford GT40, panning motion shot taken from a distance with a full car in focus, 3/4 view, high end film camera, 200mm lens, f8, 1/100s, Kodak Gold ISO 800 film --ar 16:

GT40_KodakGold.jpg
 
Last edited:
It looks like people are getting very good results with a very wide variety of prompts.

I saw someone posting an interesting image with lots of camera settings that made little sense to me. The person posting didn't understand photography at all and just took what Bard gave him.

The prompt specified a wide angle shot using a telephoto lens. It specified an ISO value, but not the camera or film used. It specified an Aperture, but didn't specify a focal length. The image looked like a tripod panning motion shot with a fast moving car and blurred background which is a challenging image to capture, but the prompt specified manual focus. The car was also at a 3/4 view, so the focus would have to change very quickly while panning the camera. Basically impossible to do manually.

I thought about what was specified for a while and the image created. Even though there was superfluous data that I'm sure Midjourney ignored, the rest could be made to work given there is no real camera.

A telephoto could get a wide shot if taken from far enough away. With a car coming at you, being far away would be a good thing. The shutter speed of 1/800s seemed very short for panning blur shot, but considering the distance and it being a telephoto lens, it might work. Also given there is no real camera Midjourney could make that shot work.

So more power for Midjourney from taking that prompt and creating a great looking image with it.

---------------------------
Below I created a more photographically accurate prompt that seems like it could work to create the image that the other guy was after. I'm not saying that I could take this shot in real life, but at least it seems possible. I also notice that at f5.6 the rear quarter panel is out of focus which makes sense. At 200mm f8 would probably be better to get the whole car in focus and because of the panning action the car would still have plenty of separation from the background.

/imagine Ford GT40, panning motion shot taken from a distance with a full car in focus, 3/4 view, high end digital camera, 200mm lens, f5.6, 1/400s --ar 16:9

View attachment 24240

I tried the exact same shot changing the prompt to f8 and the rear quarter panel is now in focus.
So I think that the camera settings do actually work and impact your images, but that Midjourney will do its best to make the most out of camera gibberish as well :)

View attachment 24241

One more thing bothered me. The wheel's hubs didn't seem blurred enough.
So I specified the car's speed and dropped the shutter speed and one of the 4 images it gave me showed the hubs blurred. Of course in this image the car is no longer presented in a 3/4 view. So still not exactly what I asked for.

/imagine Ford GT40, traveling 100mph, panning motion shot taken from a distance with a full car in focus, 3/4 view, high end digital camera, 200mm lens, f8, 1/50s --ar 16:9
View attachment 24242

It seems that Midjourney ignores ISO completely. I tried to get some film grain by specifying Kodak Gold film and saw none, however I did get the warm hues that you would expect with Kodak Gold on all 4 images. I also tried Fuji film and it had the greenish case as expected in some, but not all images.

/imagine Ford GT40, panning motion shot taken from a distance with a full car in focus, 3/4 view, high end film camera, 200mm lens, f8, 1/100s, Kodak Gold ISO 800 film --ar 16:

View attachment 24243
Wow, all of those pictures are amazing. Is this software available for free online?
 
Wow, all of those pictures are amazing. Is this software available for free online?
It is online only.
Midjourney no longer has a trial membership.
https://docs.midjourney.com/docs/plans
You use Midjourney through Discord.
First you create a new Discord Channel and invite the Midjourney bot to your channel and then it will respond to your requests.


1685466842019.png


Once you have it setup, you just type /imagine and it will say prompt and you ask it for whatever you want.

1685466915619.png
 
AI servers to drive NPC's so they react and talk more realistically.


Wow, I didn't know about their supercomputer that companies can use remotely. Thanks for that. I've bought Nvidia GPUs exclusively ever since I bought my first video card in the late 90s. It was Nvidia's competitor and ran like crap on most games. It was ATI.
 
AI servers to drive NPC's so they react and talk more realistically.
Cool, but man, that video is primarily an extended ad for Nvidia stock ;)

The ability to have their AI supercomputer tweak the logic of customer game applications to run efficiently on 'consumer' platforms is brilliant.
I'll be keeping an eye on their partnership with Mercedes self-driving applications.
 

Latest posts

Back
Top