If somebody have the necessary VectorDisplacementMap(VDM) training data, here's an idea:

"Disclaimer: Cover picture comes from the article linked in the 2nd link on the Blender Community, ticket WqWx below"

I had a discussion on comfy channel and mentioned Vector Displacement Maps. https://download.autodesk.com/global/docs/softimage2014/en_us/userguide/index.html?url=files/GUID-21452F12-9509-4F96-BFC0-F63F8DDD6465.htm,topicNumber=d30e600255
Here is a Softimage3D explanation of what VDM textures do.
https://blender.community/c/rightclickselect/WqWx/?sorting=hot
This Blender community post explains well what VDM do. It's been implemented since then but the description is great right where it's been proposed.

Training a checkpoint and a LORA, possibly compatible with controlnet but also with direct generation right out of Ksampler, on a large quantity of VDM maps could allow the enhancement of Hunyuan-3D, SPAR3D, SV3D, and Trellis which might enhance a serious work pipeline while producing material that can be considered "slighty-AI-assisted-but-mostly-manmade".

Ultimately, what 3D artists sees as the zenith of rendering is Ray-Marching environments.

Ray Marching can be achieved inside Blender3D's volumetric renderer which is fantastic.

Bringing together various techniques, we could eventually put AI to serious work with serious "I.P. material" in a serious industry by producing all the VDMs necessary to create an illusion of geometry in a rendered world without any vertex geometry whatsoever.

By training a checkpoint solely on VDM textures, as many as there can be, if you want to generate a VDM ear, it will likely produce something similar to All is needed is enough instructions so that it understands that the green,red,blue cloud on the left will produce the ear on the right.

Doing so would allow to make VDM the new standard of geometryless rendering and it would make Hunyuan renders much simpler as a lot of extra detail can be contained in 1x VDM so if you make a UDIM with VDM for each UDIM square, then you can bake a very complex character onto a very simple mesh.

With tools like https://github.com/NVlabs/CF-3DGS , it could be very possible to generate a simplified mesh but get that extra squeeze of extrusion for important details using VDM and it would make so any ray tracing graphic card could render geometryless graphics in an optimized environment and maybe it would bring solid improvement in videogames, making AI meet real artists where their real talent can be turned into a materialized shape using good workflows in a more complex pipeline involving real artists.

Imagine an image2vdm node which could both produce a texture properly unwrapped to match the vector deformation once applied to a plane, or to a box, or to a ragdoll model.

Blender Base Mesh could easily benefit from UDIM-based VDM propagation and with this method, display photorealistic renders really fast just with a few pictures from your own artwork or an actor in disguise that you hired. Your hypothetical best friend Barbara or Billy-Bob, disguised as a pirate.

Projecting VDM depth instead of simple Z-only depth maps combined with their normal maps have a serious advantage as VDM produces creases which depth maps doesn't.

Over all, not only could a character base mesh be projected out of 1 VDM but a 2nd VDM could add it the extra crisp and details to give it realism.

["A visionary"-The Crowd]

I hope you find use in this article and can share your thoughts, progress, even training data, and I even hope that we end up with something nice by putting a bit of each other's talent in it along our own.

Eventually, we could produce VDM brushes for Blender on demand and freely create without headaches about fine details. They'll come last. Just building your idea without interruption as inspiration flows.

Feels like a fair quid pro quo between artists and modern AI tools which makes so artists are empowered, not AIs.