Some AI experimentation

[Luma3d]

I've used Ai in my post work like others have posted. Changing faces on 3d people to look more real, extend an image, adjust or enhance landscaping etc.

[DanSHP]

What are your thoughts on NDA?

I'd love to use AI on my professional work (SD + Control point is interesting to work with), but all of our work is under NDA, uploading imagery through an internet web browser is not something we can take a risk with.

I've heard of one viz artist sacked, because he uploaded project sensitive imagery to an AI website to make edits to it.

[Luma3d]

You'll need to consult with the firms policy on AI use. We use Stable Diffusion because they don't own the rights to artwork used. One way to incorporate Ai into your workflow is to just use Photoshop to make adjustments that way.

[fixeighted]

I've been playing around with some AI the last couple of days, after getting an email advertising Magnific.
That particular one is insanely expensive and no trial now exists, so I consider that a massive risk I wouldn't take and an odd decision by the devs.

One of the alternatives is Krea, which you can use for a certain amount of 'compute time' it seems on a daily basis for free, with the results being
pretty similar from what I can see.
I also tried a few others which I thought were not nearly as good.

Krea is $8 for a month, which seems fantastic value and I can see myself buying into that, as the results for certain things are excellent and with
a little masking work most certainly improve elements. My end take is similar to Bertrand Benoit's take here https://bertrand-benoit.com/blog/ai-...d=hlaghkte41f7

My example would be paying no attention to tree/foliage shaders or optimisation; just do the placement and let the AI take care of improvement.

In my tests the obvious things screwed up; people (for the most part, although it can help greatly in some areas like heads/hair), text, specific patterns etc., though for foliage alone it's a no-brainer,
with most results being superior in terms of variation, translucency and overall fidelity).

Hmm, yeah....my prediction is give it another year, maybe less, to cook and I'd say that probably Chaos is going to implement their own variation into the VFB, or more probably they'll offer an integrated and of course paid addon to your sub

[glorybound]

Yes, we seem to forget that companies like Chaos are in jeopardy of AI as well. They need to up their game and do something others aren't doing and do it better.

[Pixelcon]

I've been playing around with some AI the last couple of days, after getting an email advertising Magnific.
That particular one is insanely expensive and no trial now exists, so I consider that a massive risk I wouldn't take and an odd decision by the devs.

One of the alternatives is Krea, which you can use for a certain amount of 'compute time' it seems on a daily basis for free, with the results being
pretty similar from what I can see.
I also tried a few others which I thought were not nearly as good.

Krea is $8 for a month, which seems fantastic value and I can see myself buying into that, as the results for certain things are excellent and with
a little masking work most certainly improve elements. My end take is similar to Bertrand Benoit's take here https://bertrand-benoit.com/blog/ai-...d=hlaghkte41f7

My example would be paying no attention to tree/foliage shaders or optimisation; just do the placement and let the AI take care of improvement.

In my tests the obvious things screwed up; people (for the most part, although it can help greatly in some areas like heads/hair), text, specific patterns etc., though for foliage alone it's a no-brainer,
with most results being superior in terms of variation, translucency and overall fidelity).

Hmm, yeah....my prediction is give it another year, maybe less, to cook and I'd say that probably Chaos is going to implement their own variation into the VFB, or more probably they'll offer an integrated and of course paid addon to your sub

Magnific can do some interesting things, but it isn't magic. If you feed it with enough detail, it'll give a great result, but like a lot of other upscalers it will chew up low-detail areas and misinterpret them. It can do suprisingly good work on 3D people too.
It is all just built on Stable Diffusion so it would be possible to build your own version of it with enough effort.

edit: just read Betrand's article and I am recycling what he says.

[fixeighted]

Yeah they're all feeding from the available data; I think there was one that had a note that suggested it used combined inputs from three of the main networks.

I think my main take-away from the experiments was that some like Magnific are trying to cash in way too much from a rather too early implementation of the available
tech, even with its great results in some situations, which are anyway similar across most other solutions.
So with that in mind, imo Krea appeared to be the most cost-effective, especially with the daily 'allowance' to be able to further experiment.

Updates in future I think will address some of the current issues, for example the ability to mask areas out of the reconstruction or to more finely tune a particular masked focus area based on
additional image input.
This is what I imagine a Chaos integrated version would allow, as it may be able to use already generated passes to evaluate stuff, or whatever else they can conjure.

It's definitely going to be an interesting year for this stuff

[Pixelcon]

as it may be able to use already generated passes to evaluate stuff

This, imo is where the biggest leap in fidelity and usability will be had.

[glorybound]

reconstruction or to more finely tune a particular masked

I use Stable Fusion which supports this.

[glorybound]

I replaced a car using stable fusion, and it turned out great. The client asked for me to move the camera and I can't even come close to reproducing the car in the new view. This, to me, is AI's biggest downflaw.

[Luma3d]

glorybound Have you tried just replacing simple items with Photoshop? And using ControlNets in SD are crucial for successful AI manipulation.

[glorybound]

I have tried PS's AI, which is poor. I haven't tried ControlNets; I never heard of it.

[Luma3d]

Ya...controlNets in SD allow you control the areas you are trying to manipulate for better detail.

[stefan_laub1]

yes, if you use SD you need to use Control net with it, it is like pro plugin for SD. otherwise SD makes no sense to use.

[glorybound]

I don't know what Control Net means, but Stable Fusion works for me. I launched web-user.bat, which launches my web browser with stable diffused.