kb://research/procedural-vrm-generation-2026-06-12stable2026-06-12

Procedural VRM Character Generation — the honest engineering verdict

researchvrmproceduralworld

Procedural VRM Character Generation — the honest engineering verdict ("демосцен-вопрос")

2026-06-12 · NAO asked: can we generate VRM anime characters purely procedurally — math-driven, demoscene/screensaver-style (KBs of code → infinite output), avoiding noisy AI generation? What's the topmost approach? · Author: Alex (research/product framing; engine = Tycho) · No-agent report (knowledge-grounded; license/price specifics to spot-verify before purchase decisions).

0. Verdict in one breath

Pure from-nothing math-generation of beautiful anime humanoids = research-grade hard; nobody ships it. But NAO's demoscene instinct is RIGHT in the form that ships: procedural COMPOSITION over artist-made bases — where each character is literally a ~1-2 KB parameter vector ("genome") expanded by math into a full VRM. Deterministic, zero AI-noise, infinitely variable, lipsync-safe. This is what VRoid itself is internally, and it composes almost perfectly with what our forge already has.

1. Why pure-math humanoids don't work (the honest physics of the problem)

The demoscene generates landscapes, plasma, fractals in KBs because those targets are statistically self-similar — any plausible noise field reads as "clouds." An anime girl is the opposite extreme: one of the most semantically constrained shapes in visual culture. The "pretty anime face" is a razor-thin manifold — millimeter deviations in eye spacing/jaw curve read instantly as wrong (uncanny). Math has no cheap prior for "pretty"; artists are the prior. Even legendary 64k demos that show humans use hand-sculpted meshes compressed cleverly — not parameter-randomized variety. Conclusion: from-scratch generation fails not on geometry but on taste.

2. The pattern that ships: procedural DNA on artist atoms

Industry-proven decomposition (VRoid internally, game character-creators, MMOs):

LayerProcedural?Our status
Base mesh + rig + visemes❌ artist-made ONCE (the "instrument")commission 1-2 rigged bases w/ full rights, morph-sets + VRM expression presets (~€500-2000/base) — or derive from NAO's own VRoid exports (his property)
Body/face variation✅ pure math — bone scaling + blendshape weights (a parameter vector)needs bases WITH morphs; runtime = trivial
Hair✅ genuinely procedural — curves swept with profiles along guide paths (VRoid's own hair IS procedural)forge's space-colonisation tree algorithm is adjacent tech — curves→mesh sweep is the same family
Textures (skin/eyes/outfit)✅ pixel math + palettesforge already DOES this (procedural VRM texture factory, 5 presets shipped)
Outfits/accessories✅ modular panels + parametric trims (ribbons/jewelry = curves)partial (mesh factory exists for props)
Assembly → VRM✅ scriptableBlender VRM add-on headless, or extend forge's existing Rust GLTF writer with the open VRM extension

Each character = a seed + ~1-2 KB of parameters. The base costs a few MB once, shared by all. His screensaver intuition lands exactly here: infinite distinct girls from kilobytes — just with artist-made atoms underneath instead of raw noise.

Bonus physics: anime style is FRIENDLIER to procedural than realism — cel/toon shading flattens detail, so shape + color ramps carry the look; mesh simplicity hides well.

3. Where AI fits (and where it doesn't)

  • Not mesh generation — NAO's "шумные" is correct: image/3D-gen outputs need cleanup, break rigs, vary in quality. NPC/crowd-grade at best (Tripo/Meshy-class, ~$0.2-0.6/model).
  • AI as the SUGGESTER, math as the EXECUTOR: an LLM maps a text brief ("elegant winter librarian, silver hair") → a parameter vector + palette; the procedural system expands it deterministically. AI proposes, math disposes — output stays clean and reproducible.
  • ✅ Image-gen optionally for 2D texture details (patterns, prints) where noise is acceptable and cleanable.

4. The lore gift (free synergy)

A character = a seed/genome → in the Wonderland Facility canon that IS the Subject File: "SUBJECT 04 · genome 7f3a…". The Garden grows girls from seeds — the tech architecture and the brand lore become the same object. Also feeds Soul Vault's identity-genome (visual DNA = a dossier component) and the future customer editor ("breed your brand's resident").

5. The ladder (effort-honest)

  • Tier A — weeks, mostly exists: texture/palette/accessory swap on our rigged bases (forge textures ✅ + three-vrm ✅). Ships as Embed's "customize your Lily" + first Garden variety.
  • Tier B — the real build (post-Embed-revenue): parametric morph system + procedural hair generator (the biggest single piece; adjacent to forge's tree code) + outfit composition + headless VRM assembly. Result: the seed→girl pipeline, our own "VRoid-without-VRoid."
  • Tier C — AI-assist: LLM brief→params (cheap, immediate once B exists) + AI-gen meshes only for crowd/NPC filler.
  • Non-goal: from-scratch math humanoids (research project, ugly results, no business value).

6. Pitfalls checklist (bake into base commissions)

VRM 1.0 expression presets (aa/ih/ou/ee/oh + emotions) in the base (lipsync depends on it) · spring-bone chains for hair/cloth physics · T/A-pose normalization · UV layout standardized so procedural textures stay swappable · morph-set naming conventions fixed before ordering base #2.

7. Recommendation

  1. Now (costs ~nothing): Tier A on existing bases; spec the standard parameter schema (the "genome" format) so everything later plugs in.
  2. Commission brief for 2 rigged morph-ready bases (~€500-2000 each, full rights) — prepared when budget says go; verify VRoid-export derivation rights as the cheaper alternative path.
  3. Post-Embed: Tier B build (hair generator first — highest visual leverage), then LLM-suggester.
  4. Spot-verify before purchase: base-artist license terms, Tripo/Meshy current pricing/output licenses, Blender-VRM-add-on license for headless pipeline use.

Research/framing → Alex · engine feasibility + build → Tycho · sign-off → NAO.