Some AI experimentation

[Pixelcon]

Hi there,

Just wanted to show some experimentation with AI. Interesting stuff but I can't imagine I'll be overhauling my workflow just yet. It certainly is going to be handy for replacing certain elements like vegetation and 3D people.
Using stable-diffusion and many many many different settings for each part of the image.
Took a solid day to get it from raw-render to the "finished" image (which was quite a bit longer than it took me to do the traditionally-created image in PS with lots of photobashing and matte-painting. Surely this can be sped up with more experience and I'm not at all satisfied with the result vs my original version.
I imagine when these AI interfaces can make better use of the information in the 3D scene it'll be much more powerful. It's a bit hard to wrangle the correct scale of elements it generates.

*Shameless plug* I'll be going over this in-depth on my Patreon next month.

Raw render:


My Traditional output:


AI attempt:

[oglu]

The thing with ai is. What if the client needs it ultra fast. And what if i get the same money.
At the end money wins. And thats why i have troubles right now fighting against ai in our Studio.

[Pixelcon]

The thing with ai is. What if the client needs it ultra fast. And what if i get the same money.
At the end money wins. And thats why i have troubles right now fighting against ai in our Studio.

I guess that's the age-old argument right? Quality vs. speed and pricing. If you're not actively working to position yourself as a studio/artist that creates work that is priced for its quality, then you're ineviatbly going to be drawn into the pricing war that is headed our way.

[glorybound]

It is a one-revision, one-view proposition, much like creating a rendering in Photoshop (matte painting). Thirty years ago, I used to do PS renderings, mainly elevation from 2D CAD. If the CAD were updated, I would start over. What if your clients asked for three views of your AI project? I spent about an hour replacing some overhanging trees in one scene (framing the image). Over and over, it gave me weird results. I got the left tree to work (very subtle improvement) but gave up on the right tree, which was a little more complicated.

I see a whole lot of people getting screwed, thinking they are hiring a professional to get someone who is typing in some commands and getting random results. Using it to improve an already good image would work, but for me, that is once an image has already been approved, and I am fine-tuning things that the client might not even appreciate (a little nicer overhanging trees).

I am watching NVIDIA's AI talk, and they are talking about asking their AI to build warehouse 3D models. That is cool! Upload a photo of a light fixture and get an excellent 3D model; that is my gravy. Heck, upload 2D CAD files, have an AI build the model, and fix mistakes since there are always 2D CAD mistakes.

When SketchUp came out, people said 3D modelers would be out of business since it is so easy to model in SketchUp. SketchUp is so bad that I can't use the models I am sent. If they bring the SU model to color, they look for free textures, slap them on the model, and are very low-quality. It might fill their immediate design development needs, but if they need better, they find us; it didn't kill us; it made us more valuable.

Sites like TurboSquid and 3DSKY were going to kill asset creators, but they didn't. I might glance; if I can't find it, I have someone I use to model it (much better than I am). However, if you are working on your own fun stuff and any fixture will work, go for it. Like AI, sites like TurboSquid and 3DSKY are for the hobbyist.

I do believe AI will be good enough for anyone to upload a Google Earth image, trace out the boundaries, and it'll design a structure using local building and structural codes. Real-time estimates and construction drawings will spit out in seconds. I think I will be long gone, and this is unless the AIA creates a union to fight AI.

Just a note: I used AI to check my grammar, and I have spent more time fixing what AI is changing than it was for me to type it.

Just my thoughts. Great image, by the way.

[glorybound]

This is how I use it. I replaced the landscape in the back and the car in the front.

[oglu]

If the studio head seas 60% more money for 10% less quality you know the outcome. Times do change fast. In some areas its impossible for humans to stand against ai.

[Pixelcon]

If the studio head seas 60% more money for 10% less quality you know the outcome. Times do change fast. In some areas its impossible for humans to stand against ai.

Sure. But I personally think a studio that's heading in that direction is inevitably doomed.

[oglu]

And i say it the other way around. If you dont adapt ai workflows you are doomed. Other will be faster and have more quality over time.
If ai is able to create a realistic tree in 5min. It wont help you you get a half decent tree in 3 weeks.

[Pixelcon]

And i say it the other way around. If you dont adapt ai workflows you are doomed. Other will be faster and have more quality over time.
If ai is able to create a realistic tree in 5min. It wont help you you get a half decent tree in 3 weeks.

As I said before, IMO you're entering into a pricing war you will never ever win no matter how fast you get, or how good the "quality" is you can achieve regardless of AI's involvement. If you don't fight based on price, you might have a chance of surviving.

[glorybound]

Imagine having a staff member whose sole job is to place people into scenes. Some of these people are very talented, and they pull off incredible things. I avoid people in my images because I am not good at them. Now, with AI, I can put an average-quality 3D person in my scene and quickly replace it with an AI version. The person whose job it was is no longer needed.

The same goes for trees and landscaping. I replaced all of my trees in the below image using AI in about a minute, significantly improving my image. They are not perfect, and pixel peeping reveals all, but it works for web graphics, which is what most of my clients use their images for. Maybe now I can spend more time on other things that AI can't help me with. Once these AI models are perfected, they'll be expensive to use. I heard that it takes the energy of a small city to create an AI model. You can't crack an AI engine, so it may only be available to professionals who can afford it.

By the way, I tried dozens of times to replace the road using AI, but nothing was better than the original. It's all trial and error. Even the trees, I spent some time painting out errors in the AI trees and plants.

Of course, the second image is raw out of the VFB.

[fixeighted]

I don't see any difference between the two image with regards to the foliage/trees.
Did you mean to post a comparison maybe?

[fixeighted]

Anyway, on the subject, which is of course broad and subject to literally constant change; I just did a job wherein I needed some water spurts to indicate leaking pipes.

PS generative AI gave me a load that I could use, which was a massive time saver and ofc negated having to to buy them from stock sites or create them with sims, so those
stock sites will suffer I guess, as it becomes simpler to augment our own pre-existing images.

Taking our jobs?....Hmmm, maybe a bit but I suppose we need to simply adapt and embrace our future robot overlords

The invention of the Spinning Jenny is one historical example that caused uproar and much loss of livliehood in the immediate term...

[glorybound]

I don't see any difference between the two image with regards to the foliage/trees.
Did you mean to post a comparison maybe?

It is subtle, but there is a difference. The top is the AI processed, and the bottom is RAW.

[dlparisi]

It is subtle, but there is a difference. The top is the AI processed, and the bottom is RAW.

What did you use for this?

[glorybound]

I used Stable Fusion.