Cardona Pipeline Tools / Living product evidence

MeshForge

From one image to a reviewable 3D candidate.

MeshForge runs Stable Fast 3D locally, reopens every GLB in Blender, measures the mesh, and keeps download locked until a person reviews the candidate. The workflow proves a controllable path to useful visual geometry. It does not claim CAD accuracy.

Inspect the test ladder Walk the pipeline Open the source

Runtime: Local RTX 3080 Ti
Model: Stable Fast 3D
Authority: Human export review

MeshForge Studio showing source controls, a 3D candidate viewport, topology findings, and locked review actions — Implemented local workspace. Candidate download remains withheld until the review checklist is completed.

01SourceApproved image
02GenerateOffline model
03InspectBlender evidence
04ReviewHuman judgment
05ExportExplicit approval

Working thesis

A mesh is a candidate before it is an asset.

Image-to-3D can create useful visual starting points quickly. Production value depends on what happens next: source control, reproducible settings, static inspection, visible defects, and a person who decides whether cleanup is justified.

01 / Input

Control the silhouette

Plain backgrounds and complete object bounds reduce ambiguity before inference begins.

02 / Evidence

Measure what reopened

The saved GLB is inspected after generation. The page reports that artifact, not a preview alone.

03 / Authority

Withhold the handoff

Generation success does not unlock download. Review and publication remain separate decisions.

Portfolio benchmark / 2026-10-09

Increase complexity until the model breaks.

The ladder moves from a compact manufactured form to thin aircraft structures. Each case is useful even when it fails because the same inspection contract records the failure.

Case 01 / Simple

Verification token

Tests silhouette recovery, thickness, bevels, and one recessed surface without thin appendages.

Rendered MeshForge candidate retaining the verification token silhouette and amber circular inset with visibly softened surfaces — Recognizable silhouette and inset, with softened material detail and no separable body or inset roles.

Interactive GLB artifactGLB 2.0 / 1.09 MiB

Loading interactive candidate.

Inspect the artifact: drag to orbit, scroll or use + and - to zoom, and use arrow keys after focusing the viewer. Download exposes the measured benchmark file for independent inspection; it does not change its review-required production status.

Profile: Preview
Status: Review required
Runtime: 80.76 sec
Triangles: 35,796
Components: 1 logical / 222 raw
Open edges: 0 logical / 5,040 raw

What to inspectDoes the candidate preserve the hexagonal body, shallow top inset, and visible thickness without inventing an unstable underside?

Production readingThe basic form survives, including the inset. After attribute splits are welded, the candidate is one logical manifold component. Its softened form, fused material role, and missing semantic labels still make it a visual blockout.

Token accuracy program / Pilot v4

Measure the failure, then rebuild the right layer.

Version 4 reopens the exported GLB and verifies the graphite body and amber inset as distinct PBR roles. Geometry and smoothing remain fixed while base color, roughness, metallic response, opacity, transmission, and emission become measurable.

Original Stable Fast 3D verification-token candidate with soft distorted geometry — **33.89 / 100**One logical manifold component after attribute splits are welded, one material, and no role-labeled semantic or regional-shading evidence.

V4 verification score~~33.89~~99.29+65.40 points

Tall graphite hexagonal verification token with an amber circular insert recessed below the top plane — **99.29 / 100**Two logical components, zero logical non-manifold edges, a 100 material subscore, and passing material and shading gates.

Body proportion: 0.254 H / WTaller source-matched profile
Circle coverage: 55.4%Inset diameter / hex width
Inset depth: 2.12%Below the top plane
Logical topology: 2 parts / 0 open edgesAll six pilot GLBs reopened
Normal contract: 100% coverageFlat and smooth regions passed
Material contract: 100 / 100Both named PBR roles passed

Interactive corrected artifactGLB 2.0 / 106.2 KiB

Loading the corrected v4 GLB.

Inspect the verified artifact: compare material response, wireframe topology, RGB normal direction, and false-color body and inset IDs without leaving the viewer.

Material Reconstruction v1

Four views isolate four different failure modes.

The source sampler measured median graphite #464442 and amber #a66d28. The verifier then reads PBR values after GLB reimport, while these renders expose the resulting surface response.

Canonical token under neutral studio lighting showing the graphite body and recessed amber inset — **Neutral studio**Overall body and inset appearance

Canonical token under a low grazing light that exposes bevel highlights and roughness — **Grazing light**Highlight width and bevel response

Canonical token backlit against a dark field with the amber inset still separated — **Dark silhouette**Edge and inset separation

False-color material ID render with the body in blue and inset in red — **Material ID**Blue body / red inset roles

Body: Source #464442 / M 0.48 / R 0.52
Inset: Source #a66d28 / M 0.03 / R 0.46
Color tolerance: Linear delta at most 0.08
Material gate: 100 / 100 / Passed

Six-view sheet of the corrected tall token showing its recessed amber insert from hero, front, quarter, side, and rear views — Multi-view contractOne design stays in one split.
The curriculum advances from one hero view to an opposed pair, a three-view shape set, and the complete six-view set. Views from one design never cross train, validation, and test splits.

01
Canonical dataset
Bounded geometry and PBR variations scale from the six-design pilot to 5,000 deterministic designs.
02
Semantic geometry verifier
Height, circle coverage, signed inset depth, lip width, symmetry, topology, and materials produce an explainable score.
03
Multi-view curriculum
Four stages add hidden-view evidence without leaking one design across data splits.
04
PBR recovery
Named graphite-body and amber-inset roles make material separation measurable.
05
Hard-surface reconstruction
A token-specific plan rebuilds the tall body, cuts the cavity, and lowers the insert below the top plane.
06
Normal contract
Explicit normals keep planar faces flat and bevel, cavity, and inset transitions smooth after GLB export.
07
Material verification
Reimported PBR values and four controlled renders verify role identity, color, roughness, metallic response, and opacity.

Hollow vessel program / Pilot v2

Openings, cavities, handles, clear glass, and opal glass have measurable contracts.

Eight sprints now cover three vessel classes. A ceramic mug verifies its cavity and handle. A wine glass adds a thin rotational profile and transparent optics. An original yunomi-inspired milk glass adds a gentle taper, thick lip, low foot ring, and softly translucent opal material.

Designs: 93 variants per class
Evidence: 90 views54 shape / 36 diagnostic
Score: 100.0 avg99.99 to 100.0
Gates: All passed9 reopened GLBs

Canonical case 01 / Ceramic

A mug must be hollow before its handle matters.

99.99 / 100

Warm ceramic mug benchmark with a visible open cavity and a separate curved handle — Canonical mug v1. The analyzer reopens the export and measures the opening, cavity, wall, bottom, handle clearance, and two attachment regions.

Opening / diameter: 90.59%
Wall / diameter: 4.71%
Cavity / height: 92.06%
Handle clearance: 32.51%
Logical topology: 2 parts / 0 open
Normal contract: Passed

Interactive mug artifactGLB 2.0 / 46.5 KiB

Loading the ceramic mug GLB.

Inspect the vessel: use the section cut to expose the cavity and bottom thickness, then compare normals, wireframe, and part IDs.

Side diagnostic render showing the mug handle opening and its two body attachments — **Negative space**Clearance and attachments

High-angle mug render showing its opening, interior cavity, and rim wall — **Rim section**Opening and wall evidence

False-color mug render separating the body and handle — **Part ID**Body and handle roles

Canonical case 02 / Optical

A wine glass adds thin geometry and portable optics.

100.0 / 100

Dark-field wine glass benchmark with a thin bowl, narrow stem, and broad foot — Canonical wine glass v1. The profile is measured as geometry; transmission, roughness, and index of refraction are checked as calibrated workflow targets.

Opening / bowl: 83.57%
Wall / bowl: 1.78%
Height / bowl: 2.118
Transmission: 0.9838
IOR: 1.4943
Logical topology: 3 parts / 0 open

Interactive glass artifactGLB 2.0 / 106.9 KiB

Loading the wine glass GLB.

Inspect the vessel: use the section cut to expose the thin bowl wall, then verify the bowl, stem, and foot as separate semantic parts.

Profile symmetryBowl, stem, and foot

High-angle wine glass render showing the open bowl and thin rim — **Rim section**Opening and wall evidence

False-color wine glass render separating bowl, stem, and foot — **Part ID**Three semantic roles

Canonical case 03 / Opal glass

A yunomi-inspired milk glass joins form and optical evidence.

100.0 / 100

Warm opal milk-glass cup with a gentle outward taper, thick open rim, and low foot ring — Canonical Japanese milk glass v1. This is an original benchmark form informed by yunomi proportions and the soft surface character of contemporary Japanese soda-glass cups. The analyzer measures the taper, cavity, thick lip, foot ring, partial transmission, roughness, and IOR after reimport.

Base / top: 86.69%
Opening / top: 89.74%
Wall / top: 5.13%
Transmission: 0.19159
Roughness / IOR: 0.2946 / 1.5101
Logical topology: 2 parts / 0 open

Interactive milk-glass artifactGLB 2.0 / 102.9 KiB

Loading the milk-glass GLB.

Inspect the vessel: use the section cut to expose the thick lip and deep cavity, then compare body and foot-ring IDs with the exported normals.

Profile contractTaper and low foot ring

High-angle milk-glass render showing the thick open rim and deep interior cavity — **Rim section**Opening, wall, and cavity

False-color milk-glass render separating the body and low foot ring — **Part ID**Body and foot-ring roles

08
Hollow geometry
Real openings, interior cavities, walls, and bottoms replace surface-only silhouettes.
09
Handle topology
Clearance, tube thickness, attachment span, and semantic separation become measurable.
10
Rotational glass
Bowl, stem, foot, profile symmetry, and thin walls receive independent contracts.
11
Optical PBR
Transmission, roughness, IOR, opacity, and emission survive export and reimport checks.
12
Vessel benchmark
Six GLBs, 60 views, explainable scores, and eight gates form the first repeatable pilot.
13
Milk-glass form
Taper, thick lip, cavity depth, and a low foot ring define the original yunomi-inspired profile.
14
Opal optics
Warm color, partial transmission, soft roughness, and glass-range IOR form a bounded contract.
15
Three-class regression
Nine GLBs and 90 views pass the same reopened-artifact evidence boundary.

Liquid benchmark / Pilot v2

The volume is verified. The foam has its own contract.

Six sprints add sealed coffee, red wine, and milk volumes plus a separate coffee-foam cap. Blender reopens every GLB and checks containment, fill, foam coverage, thickness, materials, topology, parts, and normals. A separate browser surface previews bounded ripples and a procedural foam texture without changing the downloadable artifact.

Designs: 93 fills per vessel
Evidence: 93 views54 shape / 39 diagnostic
Foam: 70 to 82%0.045-unit cap
Gates: All passed100.0 average

Interactive coffee and foam previewGLB 2.0 / 46.5 KiB

Loading the coffee-foam preview vessel.

Preview content Fill level68%

Try the preview: enable coffee to pause rotation and enter a fixed top-quarter view with the coffee ring and white foam visible. Use Liquid view to restore that framing after orbiting. Drag across the vessel to disturb the coffee below. The download keeps the verified 68% coffee volume and separate 76% surface-area foam cap.

High-angle render of a ceramic mug with a tan foam cap and visible dark coffee ring — **Coffee + foam**76% surface area / 0.045 cap

High-angle false-color render separating the coffee foam cap, coffee ring, mug body, and handle — **Foam part ID**4 roles / 0 open edges

Verified transparent wine glass holding a sealed red wine volume — **Red wine in glass**Transmissive PBR / 52% fill

Verified Japanese-style opal glass holding a warm milk volume — **Milk in opal glass**Soft transmission / 64% fill

16
Sealed volume
Liquid profiles derive from measured vessel cavities and export as named manifold parts.
17
Containment verifier
Fill, rim clearance, wall gap, surface planarity, and volume survive GLB reimport.
18
Liquid materials
Coffee, red wine, and milk retain separate calibrated PBR contracts.
19
Bounded ripples
A fixed-step height and velocity surface adds optional browser interaction with a static fallback.
20
Three-vessel regression
Nine GLBs and 93 views passed every reopened-artifact gate at 100/100.
21
Verified coffee foam
A separate matte cap passes coverage, thickness, contact, containment, material, topology, and part gates.

Why test a V-22 class aircraft?

Aircraft expose the hidden cost of one view.

A tiltrotor combines a broad symmetrical airframe with very thin blades, separate nacelles, deep gaps, and an underside the camera cannot see. A single photograph also freezes one mechanical configuration. Those properties make the category a useful stress test and a poor candidate for dimensional claims.

Thin structuresRotor blades can disappear, fuse, or thicken.
Repeated symmetryThe model must infer an unseen side while preserving paired parts.
Hard intersectionsWing, nacelle, and fuselage joints reveal fragmentation.
Configuration ambiguityNacelle angle and rotor position describe one moment, not a parametric mechanism.

Public-domain side view photograph of a V-22 Osprey in flight — Stress reference 04V-22 Osprey side view
Public-domain U.S. government photograph. Used to test visual reconstruction limits. It is not a source for engineering geometry, dimensions, flight simulation, or manufacturing.
Review source and attribution

Pipeline walkthrough

Every stage leaves an artifact.

The workflow separates model output from measured evidence and measured evidence from human approval.

01
Source contract
Approve one bounded object image
PNG or JPEG, 20 MB maximum, complete silhouette, plain background, and enough margin to separate the object from the frame.
Artifact
Hashed source image
Authority
Human
02
Local generation
Run the pinned model offline
Stable Fast 3D runs in a pinned GPU container with network access disabled. Profiles expose a bounded set of texture and remesh controls.
Artifact
Textured GLB
Boundary
Network none
03
Static inspection
Reopen what was saved
Blender measures vertices, triangles, connected components, boundary and non-manifold edges, UV layers, materials, and packed textures.
Artifact
inspection.json
Authority
Deterministic
04
Cleanup planning
Describe repair without changing geometry
A bounded planner converts findings into a repair sequence and estimated cleanup distance. It cannot edit the candidate or approve it.
Artifact
cleanup-plan.json
Boundary
No mutation
05
Human review
Inspect silhouette, material, and topology
The viewer supports material and wireframe modes. The reviewer records explicit acknowledgment before candidate download becomes available.
Artifact
Approval record
Scope
Download only

Implemented baseline

The first Studio run completed and stayed blocked.

The official Stable Fast 3D chair reference completed through the current interface. The output reopened successfully and retained textures. Topology findings correctly kept the job at review required.

Generation time: 87.11 secondsEnd to end through the Studio path
Candidate: 697,152 bytesTextured GLB reopened in Blender
Geometry: 12,143 vertices19,984 triangles
Topology: 167 components4,134 boundary and non-manifold edges
Materials: 1 UV layer2 packed textures, 0 missing
Decision: Review requiredDownload withheld, publication authority none

Operating boundaries

Useful visual geometry has a narrow claim.

These limits define what the current implementation can support and where a separate workflow must take over.

No CAD claim

The candidate has no authoritative dimensions, parametric history, tolerances, or manufacturing intent.

No hidden-view proof

One image cannot verify the far side, underside, internal volume, or mechanical articulation.

No automatic repair

The planner recommends work. It cannot alter geometry, merge components, or silently replace the source.

No autonomous release

Approval unlocks a candidate download only. Publication and downstream use require separate authority.

Continue inspecting

Source, context, and adjacent 3D work

Open sourceMeshForge implementationStudio, runtime scripts, inspection, tests, and documentation Research contextLocal AI LabModel evidence, promotion gates, and measured boundaries Production contextForge3DWeb visualization and annotation for approved 3D assets

Document review / Implemented October 17, 2026

Review the report as carefully as the mesh.

A recorded benchmark now becomes a one-page document with paired source and candidate images, measurements, evidence dates, and a pending decision. A local review screen checks package hashes, compares text, and binds the document handoff to the reviewed report and previews.

Walk the document pipeline