Blender Nodes need AI nodes

This are turning again in AI vs Artists discussion. This is not the topic for it. If you don’t agree with AI in Blender go elsewhere. I wish this is a topic to discuss where AI can improve Blender.
Edit: i don’t mind that anyone says that resources should not be put in AI but in a better GUI for example, or that is too early to go AI route. Those are valid but i don’t want this turned into AI vs Artists.

Yes… and you are publishing your opinion…

And this is mine… ( you don’t want to restrict my right to have one… do you ?)

If some software is to hard to use… everbody is free to use something else…

But nowone will reach any higher level of usage if everthing is just easy…

You know: nothing is ever easy… there is no law or something:

“It shall be easy…”

I couldn’t even participate in this if i didn’t learnt english to a certain level…

But…

… this is a splendid basis for any discussion…

Happy blending.

3 Likes

There are already topics like AI vs Artists to discuss what you brought here.
For me AI is a given and humans always choose the easy path to a result.
If Blender don’t go there it will be behind, so we need to make lemonade out of it.

First shot in typical -not even reaching underwhelming- Autodesk fashion

Well, you could be sure that if something is useful it will eventually come to blender, but given the opensource dynamic it will probably come later than every commercial counterparts.

Unless someone step in and develop something on his own, like grease pencil 2D animation tools.

Don’t expect something like that from BF, at least I would be very surprised if they worked on AI assistant for blender.

Let’s see how Autodesk handles it in Maya, and how artists interact with the tool. If they release it anyway !

1 Like

I wasn’t mentioning artists nor art … ( actually i also didn’t used the character combination a-r-t at all ) and i also wasn’t suggesting that the node system was perfectly fine… i’m even not against AI in general… i just wonder myself why enhancements can’t be done by people any more and mankind has to use AI for everything ( exaggerating this of course…)

Maybe i was a bit sarcastic/provocative asking why the human brain (the most advanced neural network on earth) has to be replaced… :brain::computer:
… so i don’t see the need here… (especially because nodes are already a visualized form of programming and sometimes it is easier to write code instead of working on a great bunch of spaghetti noodle nodes… :stuck_out_tongue_winking_eye: )

Also i may felt into the intended (?) trap of the title not using a “does” and a “?” …? Or we just both catched each other on the wrong foot… :wink:

Happy blending.

1 Like

No problem @Okidoki, yeah my title was not perfect. I think i stated what i wanted for this topic in my second post answering @zeauro .

I think it is probably best for the discussion trying to perceive where the AI advantages can benefit Blender.

I saw that the topic had a high risk of developing into AI vs Artists - Note that one is often more the morality of using AI.

I don’t really see the interest of serendipity in animation of an element determined by user, like a car.
In terms of animation, that does not really make sense to specify a speed, without specifying camera angle, timing of sequence.
If you want AI to be good at creating a montage of a moving car, you will have to train it to distinguish sources of advertisement, from racing, from dashcams…

That would require a specific training for a specific purpose.
That is a lot of work for a narrow niche of users.
And for creation area, that would probably be the case for every concept.

We are currently using a denoiser that is fruit of AI.
Making 32 samples render looking like a 512 samples, that is efficient and satisfying everybody.
Because that was just about computations, a pure lost of time for user and no decisive choice removed from user.
Tedious works could be good candidates.
For example, AI to create the best automatic retopology algorithm.
AI to create a muscle system rig from model specificity.

But to train AI to do such things, you will have to pay experts to create multiple cases needed, to make disappear a part of demand, that is giving jobs to them.

2 Likes

I would think that having a giant database of videos that AI can analyse of how a car moves can help. I also think if the user don’t specifies the AI will use the most common option. So if you don’t specify what side of the road you go it will go right. If the 3d car model is bigger than a 2 way road then AI assume the car go middle for example.
Yes that implies people must be employed to run AI.

Well.
On the internet, the most common case is probably a car crash from crappy cellphone videos in compilation of fails.
What is the most common, is usually bad quality, rarely satisfying.
Somebody has to triage sources for training.
In definitive, that is that person who will define what will be perceived as common case by AI.

1 Like

On this subject, I would rather the devs. not suspend all current development on Geometry Nodes (Simulation Nodes, future Rigging Nodes or whatever is next ect…) in favor of some mystical pie-in-the-sky AI node that can do everything. Considering how much R&D is needed for other companies to create such products, an indefinite suspension would likely be required.

Why would you need the rest of Blender at all, why even have Blender on your machine, can’t you just tell ChatGPT to create what you want and give you a ‘node’ with parameters to tweak? Geometry Nodes already are friendly enough for artists without throwing in the complexity of an app. like Houdini (because of the decision to abandon the initial programming-centric design).

1 Like

I just had an idea (probably not original idea) on how those diffusion models could be used in a very professional way by rendering engines like cycles or by Blender in the Compositor stage of the render.

Rather than using the SD addon that goes wild and creates a whole stable diffusion image based on say a very rough draft, which is how people use it now. How about using it in a really professional way at the compositing stage of the render to augment the rendering engine when trying to do complex things like caustics? we complete the render then in the compositor we choose “caustics” as an fx option and in the background it prompts a capable AI model to just realistically create caustics on a new layer and mix it into the scene at whatever strength we choose? I hope this gets the idea across.

I already stated several times that i am ambivalent towards nodes unless they can be still used by persons that have not a math nature. It is a bottleneck.
But i was not claiming to stop nodes. Besides AI is a valid futur option then AI should be able to learn the Blender Nodes helping the persons i made reference to above and everyone that wants a fast scene prototyping.

Well… nodes are a computer language and this is math… so the bottleneck is someone who don’t know some mathematical principles… i mean someone don’t have to have a phd in math but someone have to to understand some basic fundamentals… like causality… computability…

For example the negation of

if a then b

is NOT : if not a then not b !

Of course someone could be fine to not use math at all… ( but then for example someone can’t even check the payroll… )… but to “demand” the simplification of mathematics… well does make no sense because it’s the purification of simplicity…

To make something easier to use without knowing anything about it just makes someone a magic (or uneducated ??) user…

(Clarks third law: 1. Any sufficiently advanced technology is indistinguishable from magic.)

the idea could work since I saw chatgpt being able to make blender addons, not complex ones but I feel like it can be done.

If I understand it right youre saying that basically one gives a description, and AI tries to “fill the gaps” and use the available nodes to do what the description vaguely says…

Well I think this idea could work properly if you were somewhat experienced with blender because your description would need to be understandable according to the parameters of the AI.

In any case, I think your perception of geometry nodes is incorrect. Because its not “hard” per se… it is something that requires a separated effort and another way of thinking unusual if you cared mostly about modeling tasks…

But the nodes are few in numbers and still catchin up, improving and so on… I would consider the gemetry nodes system to be farly new for blender so its great its generating this level of expectations but I believe lots of stuff will be done in the future on it.

If you want to see something that looks “difficult” you can check out houdini.

1 Like

Well I think it would be quite complex to train an AI that can produce a node graph that works.
The first thing would be to create the training data in the first place, that doesn’t seems simple but TBH I don’t know what it takes.

And supposed it’s possible, it’s basically the same with python generated stuff. It works sometimes, but most of the time you need to know python to fix the missing / wrong bits.
So it’s probably not that useful as you might think, first because the results will be random, and you need to know how to fix the tree. And fixing someone else’s tree is much more difficult than making your own :smiley:

But the problem is more that I think :

Maybe it would be better to start with that in the first place, because I think AI in this case is a dead end at least for a bunch of time.

Sure a bit of math is needed, but much less that one might think. Most important is to practice and understand the logic. Just like 3D , if you want to be good you need to spend a few years to practice and learn all that it takes.
I suck at math too and I had to learn the basics when I was 30 to be better at python.
A bunch of years later I’m still not that good, but I can find my way in python and GN.

But I don’t see why you should use GN if that’s not your thing. regular modeling is still the way to go in many cases, and GN is a lot about tool building, these tools everyone is supposed to be able to use.

Yes that is one reasons for AI nodes in Nodes.

@sozap
The point is not GN but everything nodes EN. What i have said is much more broad.
I used Softimage XSI ICE and studied vectors and the like, for example use a cone of a light to activate scale up and down and color variations of building modules according to the music for a projection event. I had other exemples i did in ICE but that was long ago.

1 Like

Cool ! For now no-one really knows what everything nodes will look like, it’s more a concept and it will evolve by the time we get there if that appends…
Anyway to reiterate, I agree with you that if a lot of areas relies on node, something needs to be done so blender stays accessible.
But IMO an AI where you make a prompt is a poor way of solving it .

1 Like

Ironically, Autodesk killing off XSI may be a major reason why the BF today is even able to have the resources they need for big pushes like Geometry Nodes. XSI was far ahead of its time (ICE, a core capable of handling billions of polygons, insane animation systems, ect…), and it applied a lot of pressure on other companies to pour more into R&D (to the point where everyone was easily outpacing resourced-limited Blender).

Had Autodesk not thrown its weight around, the BF would still have to announce fundraisers for anything more than incremental improvements (as they might otherwise just have Campbell and Brecht as paid developers with Ton writing code still). Geometry Nodes? Unless a volunteer who really wanted node-driven proceduralism and power stepped up, keep dreaming.

I think Autodesk killing Softimage was an event that show everyone that your work tool can just disappear even if it was at top of popularity and since it is proprietary nothing can be done. I’ll put that , the start of renting by Autodesk/Maxon and the Blender 2.8 GUI fix the 3 events that propelled Blender to be in widespread use at present.

This is what the most advanced AI available to the public can do with Python.

That is with a mega-million-dollar application using the full API. As I mentioned before, an AI node will probably be a huge time sink that would take resources away from the other objectives of Everything Nodes (in which we do not have to speculate on whether they will unlock a lot of new and powerful workflows).

2 Likes