Model Fidelity — Full Pipeline

How a forged Void Horizon character becomes a clean, riggable, mo-cap-driven model: take a known-excellent open-source base (CC0, commercially usable), decompose it into parts, and apply our forged identity onto those parts — scaled to fit — so the mesh that actually deforms is a proven, animation-grade body, not the raw forge output. Every stage is measured, not eyeballed.

The pipeline, stage by stage

Source — a known-excellent base stable

A MakeHuman CC0 body (14,738 verts, uniform UV) plus a commercially-usable mo-cap library. The base is anatomically clean and rigs/retargets reliably — the properties the raw forge mesh lacks.

assets: MPFB2 v2.0.17 MakeHuman CC0 · mo-cap retarget sidecars

Decompose & map parts stable

The reference is split into parts (body, head, hands). Our forged geometry supplies identity (silhouette + texture); the CC0 golden mesh supplies the deforming body + skeleton. Parts are scaled to align the forged proportions onto the golden topology — identity is carried entirely by the texture, so one clean mesh deforms for everyone.

Path B — golden body (task-590)

Golden body + bulk/size stable

One shared golden mesh, morphed per character for stature and build (small/medium/large × lean→heavy) as a fast macro-morph — no re-forge, no GPU.

art-gen/golden_rig.py --bulk · bulk & height tuner →

Transplant the forged texture stable

The forged per-character skin is baked onto the golden body's UV atlas, so the proven mesh wears the character's identity.

art-gen/apply_texture.py · golden bake + verify

Better-resolution face — head-map verified · face

The hi-res forged head is cloned into the body atlas with a Poisson seamless-clone: boundary tone matches by construction, so the face-to-scalp seam disappears — and, decisively, the real eyes, brows, nose and lips are restored instead of the old moment-match's featureless pale mask with empty black eye-holes. Measured on ariel_sorin: seam tone-step ΔE 5.0 → 2.5 (halved), facial-feature pixels up 4.3×. (The pale-face and hair-through-scalp code paths didn't trigger on this character — no blowout existed to pull down — so those are situational, not the headline win.)

head map before vs after: old washed-out pale face vs new with restored eyes/nose/lips — Head map + on-model, OLD (moment-match) vs NEW (Poisson clone): the featureless pale mask + black eye-holes becomes a real face. Seam ΔE 5.0→2.5.

art-gen/uv_headmap.py (cv2 seamless-clone) — committed & verified (face-fidelity)

Eyes shipped cast-wide

Four measured bugs fixed — ball-center, oversized eyeballs, bug-eyed flat faces (per-character eyelid-surface recession), and cache staleness — plus per-character/species iris colors.

art-gen/eye_refine.py · eye refinement study →

Rig + mo-cap retarget ⚠ skin weights failing gate

The golden skeleton drives the mesh; mo-cap clips retarget onto it via per-clip sidecars, producing the required animation set (23 humanoid motions: locomotion, melee, ranged, reactions).

163-bone golden rig · *_mocap.retarget.json · processed cast → · motion lab — 23 required motions live →

Measure — the acceptance gate measuring

"Perfected" is not a vibe. Every clip is run through the fixed surface-deformation harness — per-region stretch, shear, inversion, and Green strain, a pass/fail VERDICT gate, and a shred_index. The rest pose self-tests to ~1.0 to prove the harness is calibrated. A model ships only when it clears the gate under animation.

art-gen/surface_metrics.py · deformation measurements →

Live measurements — current run

Measured just now with the fixed harness. The rest-pose self-test reads a clean 1.000 (0 strain, 0 flips) — the harness is calibrated, so every number below is real deformation, not an artifact. Result: 0 of 9 clips clear the gate. The eyes and face texture are solved; the skin weights are not — the golden body tears at the shoulder and self-intersects at the pelvis under motion. This is the next fix, and it is measured, not assumed.

measurement	result
Rest-pose self-test (calibration, must ≈ 1.0)	1.000 — clean (0 strain, 0 flips)
Clips passing the deformation gate (p95 strain ≤ 0.50)	0 / 9
Worst region across all clips	shoulder_L (deltoid) — strain p95 1.914 on dodge, 3.8× gate; + inverted tris & self-int
2nd hot spot	pelvis / inner-thigh — self-intersects on every locomotion clip
shred_index range (idle → run)	2.0% → 16.6% (scales with motion intensity)
Head-map seam ΔE — before → after (ariel_sorin)	5.0 → 2.5 (halved); facial features +4.3×

Next: reweight / corrective-fix the shoulder deltoid, then the pelvis/inner-thigh, and re-run this gate until clips pass. Fix path uses corrective_blendshape.py / blender_corrective_rig.py.

Composition pipeline — data-driven mechs & walkers new · live in gameplay

The humanoid pipeline above proved the lesson: don't rig a raw forge — it self-intersects and won't deform. The mech/walker track applies the same lesson a different way. Instead of one baked mesh, a character is a recipe: a known-good CC0 golden body + grafted identity parts + a material preset + equipment slots, which the engine assembles at load time. This is now live in gameplay — the player avatar and the body-picker read the manifest and compose the model, not a pre-baked glb.

The golden-body family

humanoid

MakeHuman (CC0)

14,738-vert body, uniform UV, 163-bone rig. Skin/graft a forged identity onto it — the stages above.

mech / walker

Quaternius (CC0)

Animated Mech Pack — Stan (biped), Mike (heavy), George (chicken-walker). 43-bone biped rig, 18+ baked clips.

creature

creature-composition

The same recipe for non-humanoid frames — graft signature shells onto a riggable creature base.

Each is a known-good, riggable base we skin identity onto — the opposite of a raw forge that won't rig.

Manifest — the recipe data-driven

Per-character record: base_model + composed_parts (id → bone mount + xform) + material_preset + equipment_slots. The composition lanes emit this manifest instead of baking a glb — mount_rigid_auto derives each part's mount & transform.

species_data.json composition block · gear_compose.mount_rigid_auto

Assembler — engine builds the node tree runtime

On load the Godot assembler instances the base_model (its Skeleton3D + baked anims), BoneAttachment3Ds each grafted part at its mount (mirrored for .L/.R), applies the material preset to semantic slots, and fills equipment slots from the loadout. One assembler drives NPCs, the player, and the customizer preview — same code, same data.

Godot runtime composer · rigid bone-mount (identity + gear), skinned garments to skeleton

In-game live: avatar + body picker

The composed mech is assembled at runtime in gameplay — player avatar and the body picker read the recipe. Made editable, the same manifest is the character customizer (mix & match base + parts + colors + equipment on a live render doll).

player avatar · body picker · manifest → customizer

a01 — mis-forged humanoid → composed mech

a01 is canonically a salvage-mech, but it had been mis-forged onto the humanoid body — so its forged mech-face smeared across a human head (below left). Redone as a composed mech: Quaternius Mike base + grafted shoulder domes / chest plate / claw hands + a salvage recolor. 🤖 How close can composition alone get to the concept? — the convergence study → (six iterations toward the concept art, with the method's measured limits).

a01 in the character browser: a forged mech texture smeared over a humanoid head, garbled yellow forehead and empty face — The salvage-mech forged onto a humanoid body: a garbled face, no clean silhouette, won't rig as a mech.

a01 rebuilt as a composed mech: Quaternius Mike golden body with grafted domes, chest plate and claws, salvage recolor — Mike golden body + grafted domes/chest/claws + salvage recolor — a real mech silhouette that rides the rig.

a01 composition steps: concept art, prior flat-recolor PoC, then composed idle/walk/run/punch/kick riding the rig, plus a cannon graft — Composition, step by step: concept art (identity target) → prior flat-recolor PoC → grafted domes/claws onto Mike → the grafts **ride every baked clip** (idle / walk / run / punch / kick), and a cannon grafts onto the claw for Shoot.

a01 composed mech wielding a cannon in idle, walk and shoot poses, with a rocket pack augment — Weapons & augments are the same composition tech, one layer up: a01 wields a **cannon** (grip → hand bone) and rides a **rocket pack** — attached last, riding the animations, never baked in.

three composed walkers: r01 navy/yellow combat bot on Mike, r02 white/teal medic on Stan, spurs brass drone on George — The walker lineup — same golden bases, different recipes: **r01** navy/steel combat (Mike), **r02** off-white + teal medic (Stan), **spurs** brass/gold scout drone (George chicken-walker). Recolor + parts only.

Quaternius Animated Mech Pack CC0 validation: Stan, Mike, George, Leela mech bases with tri/bone counts and a shared walk animation — The CC0 golden mech bases we compose onto — **Stan** (6,190 tri / 43 bones), **Mike** (heavy), **George** (chicken-walker), Leela (armless, reject) — sharing a universal biped rig and 18+ baked clips.

Update (Sep 5) — composition superseded for r01 / r02 / spurs. The Quaternius-composed walker lineup above (r01 on Mike, r02 on Stan, spurs on George) has been replaced by the coherent single-mesh forge below — grafting forged fragments onto a low-poly Quaternius base read as fragmented up close; a full coherent forged mesh does not. See the current versions in the next section.

Hard-surface single-mesh forge — the mech-overhaul new · task-748

The composition method above still stands for the golden-body humanoid/organic track (skin identity onto a proven deforming body). For rigid hard-surface combat robots it hit a ceiling: bolting forged armor fragments onto a low-poly Quaternius base reads as a fragmented, low-detail bot up close. The fix is the same lesson applied differently — forge the WHOLE robot as one coherent high-poly textured mesh (the welded plates ARE the geometry, not separate grafted parts), then rig it onto a clean biped skeleton with rigid per-plate skinning (no organic deformation needed for rigid armor).

Concept driver (provider-selected) 1-2 candidates, no sweep

A single clean plain-background full-body A-pose concept per character — the armor plating IS baked into the driver image (the deliberate exception to the base-body rule: a mech's armor is the model). Codex/GPT agents use the directly exposed OpenAI image-generation tool or gpt-image-2 API; agy is a secondary remote provider, while FLUX/SD3.5/SDXL remain local fallbacks. Every cached driver records provider, image model, prompt, input hashes, output hash, and selection.

Image → 3D, image-driven no humanoid rig

Hunyuan3D-2 shape (octree 512) → Hunyuan3D-Paint → normalize (height 1.8, feet at 0) → one coherent textured master mesh, ~1-1.3M source tris.

art-gen/forge_character.sh · HY_OCTREE=512

Repair + rig static accepted · motion gated

Before the unchanged decimation collapse, identical exported positions are reconnected. That fixes the export-split/decimation fractures on the verified non-Warhound static candidates without distance welding, smoothing, or deleting geometry. The old custom 17-bone animation still shears at the thighs; the richer Rigify path must clear sole contact, posture, transition, and export-fidelity gates before it replaces any animated model.

exact-position reconnect → COLLAPSE · static surface ≠ accepted animation

Store + promote binaries LFS + R2

Forged sources and game-loaded GLBs are versioned through the private R2-backed Git LFS endpoint. A curated web model is copied to public R2 by scripts/publish-web-models.ts and served through the same-origin /models proxy. Heavy review media also goes to public R2; only small thumbnails remain ordinary Pages Git blobs.

a01 — three generations of the same character

a01 shows the whole arc: mis-forged onto the humanoid body (Aug 31, above) → composed onto a Quaternius base (Aug 31, also now superseded) → this coherent forge (Sep 5, current). Clean, no seam between "parts" because there are no parts — one mesh.

a01 reforged as one coherent hard-surface mesh: gunmetal-grey and rust-orange salvage mech, front, rear, and side views, clean silhouette with no visible part seams — a01, current: one coherent forged mesh (front / rear / side). No grafted-part seams — the armor plates are the geometry.

r01 — the walker-lineup character, redone

r01 as a composed low-poly combat bot on a Quaternius Mike base, navy and yellow — Aug 31: r01 as a Quaternius-Mike composition (navy/yellow) — the walker-lineup shot above.

r01 reforged as one coherent hard-surface heavy combat robot, chunky gunmetal-grey armor plates with amber accents, front, rear, side — Sep 5: r01 as one coherent hard-surface mesh — chunkier, higher-detail armor plates.

historical r01 custom-rig walk cycle shown across five frames — Historical r01 custom-rig walk clip, 5 frames. This contact sheet predates the current Rigify work and does not pass the motion bar: thigh shear and the weak stride still require replacement backed by deformation and contact measurements.

More of the 23-strong coherent cast

23 characters have gone through this forge so far: a01, a02, aegis_7, android_king, aura_net, chronosync, daemon_enforcer, dharma_unit_7, doc_holliday_7, echo_chrome, nexus_prime, null_pointer, r01, r02_service_bot, rev9_preacher, shadow_null_bruiser, shadow_signal_jammer, spurs_scout_drone, synthetix_primus, the_synthetic_analyst, vex_7, vite_01, zig_combat_android.

aegis_7 coherent hard-surface jaeger-style mech, dark grey plating with teal accents, front, rear, side — aegis_7 — a jaeger-style heavy frame, same method: one coherent mesh, rigid-skinned.

historical shadow_null_bruiser render before the cast-wide albedo despeckle proof — shadow_null_bruiser — this historical source render shows the original paint noise. The cast-wide med5 albedo despeckle proof removes the pinpricks while preserving panel lines and accents; the public model still needs promotion from that repaired texture path.

Current acceptance gaps:

Texture repair exists but is not yet the public source — the cast-wide med5 proof removes albedo speckle without changing mesh, skin, or accessors. The old contact sheets and public GLBs must be replaced from that reviewed output rather than described as the current result.
spurs_scout_drone needs a redo — the salvaged pre-OOM forge produced a melted, mis-posed crouch, not a clean biped; it is not shown above as a positive example and should not be treated as representative of the method.
Surface and motion promotion are separate — the Mech745 static pilot is hash-locked with zero connected-surface boundaries. The old animated assets stay labeled until the richer rig clears the live movement gates; then the accepted GLB is promoted through public R2.
Not pilotable, no cockpit yet — task-606 (pilotable walkers) and task-607 (cockpit interiors + entry/exit animations) are still pending in the queue. What's shown here is a forged, rigged, locomotion-capable body — not yet an enterable vehicle.

Part of the model-fidelity effort. Pushed to the preview branch for on-the-go review — steer any stage and it re-runs. · 🤖 a01 convergence study → · ← back to console