There is MLops as a start. A nativ integration of AI system sounds intressting but how the integration should look like i have not a clue.


https://github.com/Bismuth-Consultancy-BV/MLOPs [github.com]

Any news from the fmx presentation regarding the new feather system in H20?

Has anyone been to the houdini hive conference? Tell us something news about houdini20!

Soothsayer
Think bigger! An AI context where we can noodle up our own training models!

Maybe thinking Simpler is bigger. Houdini is too complex at times AI can help simplify it not sure people are really that interested in using their own training models if a pre-train version would already exist. Most Artist want to create art without the technical hassle only a few will go the extra mile of deep exploration. So to maintain market share or grow it Houdini will have to get Simpler
that is more of a threat than AI.

dreamblk

Soothsayer
Think bigger! An AI context where we can noodle up our own training models!

Maybe thinking Simpler is bigger. Houdini is too complex at times AI can help simplify it not sure people are really that interested in using their own training models if a pre-train version would already exist. Most Artist want to create art without the technical hassle only a few will go the extra mile of deep exploration. So to maintain market share or grow it Houdini will have to get Simpler
that is more of a threat than AI.

I think that people who are in the target audience of sideFX defiantly would prefer an AI context where we can build ai`s over having some sort of preset generator all though if possable we i`d prefer Towlie`s solution.

you could in essence already use houdini for said purpose in tops, as they have demonstrated in one of their video’s a while back. Although, you would still have to do the full coding on all of your models. It would be great if there was indeed some sort of context where we could build models with nodes which we could then use in tops or somewhere else. I do think that it is an incredibly complex technical and design challenge, and I'm willing to bet finding and hiring people with the necessary skillset to create something like this is very hard and probably also very expensive.

Why do you need a specific AI context tho. Why not just TOP + some AI-related nodes? I'm pretty sure you'll need to call some external python code in the process, so just TOP.

It's already happening. AI tool that uses text prompts for Houdini.

https://80.lv/articles/shiba-a-wip-ai-that-can-analyze-houdini-files-and-suggest-improvements/ [80.lv]

luoqiulin
Has anyone been to the houdini hive conference? Tell us something news about houdini20!

sidefx said that they showed houdini 20 feather preview in 27 april ,
soo they are going to publish a video ?

habernir

luoqiulin
Has anyone been to the houdini hive conference? Tell us something news about houdini20!

sidefx said that they showed houdini 20 feather preview in 27 april ,
soo they are going to publish a video ?

Yes, now on their YouTube channel

mandrake0
There is MLops as a start. A nativ integration of AI system sounds intressting but how the integration should look like i have not a clue.


https://github.com/Bismuth-Consultancy-BV/MLOPs [github.com]

Stassius implementation looks much more accurate:

https://github.com/stassius/StableHoudini [github.com]

https://youtu.be/Un_b8cvzxcw [youtu.be]

mandrake0
A nativ integration of AI system sounds intressting but how the integration should look like i have not a clue.

Sound propagation can be pathtraced in a nice, deterministic and realistic manner. "Two Minute Papers" YT channel mentioned a whitepaper addressing this [yewtu.be] back in 2016.
I don't see how purely AI-based models could be useful here.

mandrake didnt refer to actual sound, but the idea of having it. (as in, having X feature sounds interesting)

anyway, i'd like to be more clear about ML implementations i personally wished for in H: what i meant was ML-accelerated up-rezzing of simulations, e.g. give the system 0.1 voxel pyro sim, and it will figure out 0.01 result without actual re-simming. not this prompt mumbo jumbo

Neural Physics for realistic hair simulations (Nvidia)
damn.. i cant wait when AI will be avaible in Houdini. Imagine that you can easilly train your own models for all CFX tasks. from hair simulation thru muscle deformation and cloth stuff. Animators would love to work with such realtime feedback from neural effects.

https://blogs.nvidia.com/blog/2023/05/02/graphics-research-advances-generative-ai-next-frontier/?=&linkId=100000200652087 [blogs.nvidia.com]

Neural morph. Neural deformer presented in UE 5.2. Fully anatomically correct body deformations including muscle tension, skin sliding, etc. ALL RUNNING IN REAL TIME :-) BTW those models were trained on simulations made in Houdini! so id love to have such workflow inside Houdini ASAP :-)

I imagine, that for start, there should be 3 nodes (a little bit like Copycat node in Nuke with database of already pretrained stuff ready for download from "CATerry" website)

1st node for formating and filtering data and making it usable for training modes. 2nd node for training inside Houdini (with option for easy training on the cloud - (better cloud integration))

3rd node for reading different models. Those that were trained in Houdini but also other by 3rd party and avaible in site similar to CATerry with library of models made by SideFX and community. Of course options like adding embeddings to model, prunning, mixing differrn models together, converting etc. All this feautures should be avaible right away.

Houdini AI nodes should read all kind of data, from voxel color and velocity to particle vector movement, textures, UVs, VIDEOS! (references) vertex x,y,z position, edges. kineFX skeleton!, just EWERYTHING, including how nodes are connected together in editor :-) so we could train everything. Every type of data against every type of data and make our work easier and faster.

Here's short review of this Neural deformation in UE5 in his YT channel:

Any date for an upcoming "Houdini 20 sneak peek" ?

99% H20 event will be held sometime mid summer 1-1.5 month from now, it happens when changelog approaches 700, think sneakpeak needs one more month at minimum.

What can we expect from Houdini's 20 new VIEWPORT technology and how it'll be called? (Directx 12, Metal or something else?.. i forgot xd)?

Is there gonna be node based REAL TIME Viewport Compositing like it is in Blender right now? (its preatty cool btw!)

What else can we expect from this new technology?
.

Hope there is new implementations like Intel® Open PGL to speed up the render time in Karma.

Vray start using it in the latest build.

https://www.chaos.com/blog/path-guiding-with-intel-open-pgl-in-v-ray [www.chaos.com]

https://youtu.be/zKdZCZZbMRY [youtu.be]

I hope we can get a camera composition grid overlay for rule of thirds, Fibonacci spirals and diagonal lines. C4D has this and it makes scene block outs so much easier.

matsnyman
I hope we can get a camera composition grid overlay for rule of thirds, Fibonacci spirals and diagonal lines. C4D has this and it makes scene block outs so much easier.

Hi !! You can setup this in houdini if you want.Or
There is a HDA by Dyna studio "Houdini HDA | dy Composition Guides"
Cheers