[x3d-public] Fwd: [x3d] MSF Plenary #18 Slide Set

[John Carlson]

Ah, I was looking the DGX Spark, which is a little pricier.  If someone
needs help setting up Apache or a website, I can help.  I’ve set up Docker
images.  I’m interested in a shared resource in general, since running just
one app cost $45 for two weeks—really pricey.  My ordinary shared web
hosting costs < $20 per month.  But no node or deno runs on it.  I can
probably chip in $200 per year max.  I have no issues with setting stuff
up, but administration beyond set up and installation (backups) I’m not too
familiar with.

On Thu, Nov 13, 2025 at 5:50 PM GPU Group <gpugroup at gmail.com> wrote:

> John,
> > 1.  Who is going to pay for AI training?
> - great question.
> There are little computers Nvidia Jetson Orin Nano Developer (check
> Amazon) for 400 to 1500 $
> - they are loaded with Nvidia's AI stack for a peripheraly AI, linux OS,
> and various 3rd party have 'foundation models' for various things
> - One option for funding: there could be a little club, where those
> who want to provide the tool to their users, would chip in (note 1)
> - the data sets - web3d.org has lots of data sets / x3d example scenes,
> and the images for them can be generated with an x3d-browser (note 2)
> > 2. Is there a foundation model we can use?
> - great question.
> - presumably the Nvidia stack comes with some basics, and 3rd parties
> might have some refined models
> - AlphaFold by DeepMind: AI used to solve protein folding, used the 2 step
> confirmation method, where the AI would give a best guess, then a chemical
> bond checking algorithm would confirm or reject its logic and feed that
> back for refining the AI training, (I read somewhere, so it must be
> possible to do that.)
>
> Note 1: there could also be a shared 'resource pull tool' for club
> members, and public downloaders would see a screen with
> tool-or-resource-name [Install / Uninstall / Update / Discontinued ]
> button for each tool or resource. So someone downloading freewrl, would get
> the pull tool that allows pulling sunrise, x3dom server, castle model
> viewer, x3d-edit, ....., and the Update button would be lit when there's a
> new version.
> Note 2: complex shapes like HAnimHumanoid would be easier to train AI for
> if instead of training against a mesh, it was training against simplified
> scalar parameters, like MakeHuman / MPFB extension for Blender. There could
> be an abstract x3d type HanimGenerator, and a concrete type
> HAnimHumanoidGenerated that derives from it and HAnimHumanoid, and on
> rendering, it generates the skin Shape and textures, and coveroids /
> clothing based on parameters. Since AI trains on the scenefile --not the
> runtime scenegraph-- it would only need to learn a few dozen slidebar
> parameters and control them to generate a humanoid that looks like the
> training image. This would allow x3d authoring tools to tinker with the
> humanoid looks with slidebars, and allow x3d-browser events to change the
> look of the Humanoid on-the-fly / at runtime.
> -Doug
>
> On Thu, Nov 13, 2025 at 3:43 PM John Carlson <yottzumm at gmail.com> wrote:
>
>> That sounds good, Doug.  Questions
>>
>> 1. Who is going to pay for AI training?
>> 2. Is there a foundation model we can use?
>>
>> On Thu, Nov 13, 2025 at 1:47 PM GPU Group via x3d-public <
>> x3d-public at web3d.org> wrote:
>>
>>> "AI assisted content creation"
>>> Suggestion for web3d: train AI to generate x3d content:
>>> a) from photos
>>> b) from textual description
>>> Since chatGPT can generate images from textual description, its image
>>> output (b) could be combined with x3d-specific (a) photo to scenefile
>>> generation:
>>> textual scene description > chatGPT > images > select image > feed image
>>> to (a)
>>> To train for x3d-specific (a) a 2-step process is used:
>>> 1. pairs of (input,output) = (x3d-browser-rendered image, x3d scenefile)
>>> are fed to the AI training system to train it
>>> 2. then AI used with input image to generate a scenefile, and the
>>> resulting scenefile is checked using an automated tool to confirm or reject
>>> on xml and web3d spec conformance, and the result used to reinforce /
>>> correct train the AI further
>>> -Doug
>>>
>>>
>>> On Thu, Nov 13, 2025 at 11:36 AM Joe D Williams <joedwil at earthlink.net>
>>> wrote:
>>>
>>>> > MSF plenary, can you point everyone to slides?
>>>>
>>>> Here is the slide set for the MSF Plenary November 12 2025
>>>>
>>>>
>>>>
>>>> https://docs.google.com/presentation/d/1-zOzU38yRLasbth0JSd_9PqtJEjEiumTtLOj1q_tw7g
>>>>
>>>> _______________________________________________
>>>> x3d mailing list
>>>> x3d at web3d.org
>>>> http://web3d.org/mailman/listinfo/x3d_web3d.org
>>>>
>>> _______________________________________________
>>> x3d-public mailing list
>>> x3d-public at web3d.org
>>> http://web3d.org/mailman/listinfo/x3d-public_web3d.org
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://web3d.org/pipermail/x3d-public_web3d.org/attachments/20251113/97071bf2/attachment-0001.html>