Hi.

Vray and others are jumping on the whole text prompt texture and material generation bandwagon. This could be good For Karma at some point.

https://www.youtube.com/watch?v=MBSJRAWI6zk [www.youtube.com]

Best

Seem to be jumping into the 'gimmick' side of AI gen with both feet I see.

Voice generation ? Shitty looking watercolor filters ? 'set extension' with random, impossible to control nonsense ? Really ? This is what the guys making/using VRay see as the future of their render engine ?

Mark Wallman
Hi.

Vray and others are jumping on the whole text prompt texture and material generation bandwagon. This could be good For Karma at some point.

https://www.youtube.com/watch?v=MBSJRAWI6zk [www.youtube.com]

Best

The general attitude here is very against this kind of thing, so don't expect that in Houdini.

That being said, I'd like to see SideFX integrate some existing open source AI tools. The ONNX Inference SOP is a good step. That would be great if we can use Houdini instead of ComfyUI for Stable Diffusion workflow.

I do not think that people are against this, just against big company doing it for you. All of this can already be done with Houdini. There is no "value add" to have these, but building these DIY in Houdini could be valuable.

raincole
That being said, I'd like to see SideFX integrate some existing open source AI tools. The ONNX Inference SOP is a good step. That would be great if we can use Houdini instead of ComfyUI for Stable Diffusion workflow.

ComfyUI and Stable Diffusion are cringy and problematic for many relevant reasons.

I think that physically-based generative textures are important but for the Houdini audience, it should be a tool with intentional creativity and/or physical accuracy. It would be good to have Differentiable Rendering in Houdini as soon as possible. This lack of Differentiability is holding the community back from contributing real content at the same par as these large corporations have generating appropriated content.

Karma should be Differentiable or will be obsolete by the end of this year. It can start with Stochastic Gradient Descent:

Transforming a Non-Differentiable Rasterizer into a
Differentiable One with Stochastic Gradient Estimation [arxiv.org]



but should be as involved as Mitsuba 3 with full-spectrum and scene parameter derivatives. This should be accompanied by a Differentiable Shader Language similar to Slang and SlangTorch.






With Houdini you should not need to use those models trained on other's works. Generate your own dataset and train! If you are ok with the ethics of ComfyUI and SD then check out MLOPS.

AI is arriving in a big way but not like what you have seen by large companies. Creatives are just getting started.

The room decorator is simply a prompt to a language model that has been fine-tuned and provided a library of items to semantically search" like shown before by Nvidia:


and
https://github.com/siva-nagendra/usdchat [github.com]
and
https://github.com/GPUOpen-LibrariesAndSDKs/RenderStudioKit/blob/main/Sources/Schema/AiTools.cpp [github.com]

Nothing special from Vray here. They need to do something new because physics-based neural networks will make rendering trivial over the coming years. I do not think their business is sustainable anymore.

This is where Houdini should pay attention and the MPM solver should have been at this level as all of this is MPM-based:

PHENOMDESIGN

These are great! Very interesting. I'm doing preliminary work on a master's thesis right now and I'm pondering how differentiable rendering could be made more accessible and useful.