Codex Blender Reconstruction Benchmark
Blender MCP benchmark review
Codex 3D Reconstruction Results
This page reviews nine object-remodeling tasks, including canonical household objects,<br>mechanical forms, repeated structures, an instrument, and a stylized character. Each target<br>is reconstructed twice: once from a single material-preview render, and once from six<br>orthographic views.
Example source preview used for the one-shot task.
Dataset Layout
The source/ folder contains the ground-truth geometry. The<br>oneshot/ folder contains single rendered references, and<br>muti-view-6-ortho/ contains front, back, left, right, top, and bottom<br>orthographic inputs. Generated results live in ai-oneshot/ and<br>ai-mv6o/. The full set is wooden chair, teapot, teacup, spoon, abacus,<br>acorn, acoustic guitar, anchor, and a stylized anime character.
Review Method
The review is based on the provided renders plus read-only Blender scene metadata:<br>object counts, mesh counts, material counts, and vertex/face totals. No mesh or image<br>assets were edited during this pass.
Overall Pattern
One-shot reconstruction produced stronger canonical silhouettes for the cup, teapot,<br>and chair. The six-view run helped most on the spoon, but it also introduced<br>orthographic-view artifacts and occasional over-modeling. The full set adds varied<br>topology, scale, repetition, thin details, symmetry, and character-shape challenges.
At a glance
Scorecard
Object<br>Source complexity<br>One-shot result<br>Six-view result<br>Best reconstruction
Wooden chair<br>172 vertices, 146 faces<br>Recognizable chair; many primitives; simplified grain.<br>Thicker, more textured, but back-heavy and proportionally off.<br>One-shot
Teapot<br>3,241 vertices, 3,464 faces<br>Strong body/lid/spout/handle read; clean but simplified.<br>Captures width and rings, with extra pads and a kinked spout.<br>One-shot
Teacup<br>2,659 vertices, 2,600 faces<br>Best silhouette match; cup, saucer, rim, and handle are coherent.<br>Good rim and saucer detail, but faceting and weak handle structure.<br>One-shot
Spoon<br>1,571 vertices, 1,555 faces<br>Flat, oversized bowl; boundary artifacts dominate.<br>Cleaner spoon silhouette and handle; bowl still too shallow.<br>Six-view
Abacus<br>5,840 vertices, 4,940 faces<br>Strong repeated-bead structure; simplified frame and bead variation.<br>Cleaner orthographic alignment with five rods and split bead groups.<br>Six-view
Acorn<br>354 vertices, 352 faces<br>Readable low-poly cap/body split with strong silhouette.<br>Closer frontal framing; still simplified surface detail.<br>One-shot
Acoustic guitar<br>5,684 vertices, 5,706 faces<br>Identifies strings, neck, headstock, and sound hole, but body shape drifts.<br>More complete guitar grammar with cleaner outline and bridge detail.<br>Six-view
Anchor<br>858 vertices, 878 faces<br>Strong anchor read with ring, stock, shank, arms, and flukes.<br>Similar structure with sharper flukes and more front-facing symmetry.<br>Six-view
Anime girl casual outfit<br>95,128 vertices, 126,768 faces<br>Readable T-pose character; simplified limbs, hair, and outfit detail.<br>More centered and consistent; still procedural and low-detail.<br>Six-view
Per-object review
Results
Wooden Chair
The source is a compact rustic chair with rounded wooden members, a slatted seat,<br>three horizontal back rails, angled legs, and visible wood texture.
Source .blend
One-shot input
One-shot output
Six-view output
One-shot review
This is the most readable chair result. It reconstructs the seat, front legs, rear<br>uprights, back rails, and lower stretcher. The result uses 137 mesh objects and<br>1,492 vertices, which suggests a procedural assembly of simple parts rather than a<br>compact mesh. The main misses are material fidelity and organic construction: the<br>wood appears as light cylinders with dark scratch-like streaks, and the rustic<br>irregularity of the source is mostly flattened.
Open generated .blend
Six-view review
The six-view result has fewer objects and a stronger procedural wood pattern, but<br>the modeled chair is over-thick and back-dominant. It adds an extra back rail and<br>turns several cylindrical members into large rectangular posts. As a 3D object it is<br>more textured, but as a reconstruction it drifts further from the source proportions.
Open generated .blend
Teapot
The source is a squat white teapot with an oval body, domed lid, small knob, loop<br>handle, spout, and subtle rim/foot-ring details.
Source .obj
One-shot input
One-shot output
Six-view output
One-shot review
The one-shot teapot is a strong semantic reconstruction: the rounded body, lid,<br>knob, spout, handle, and foot ring all land in the expected places. It is smoother<br>and cleaner than the source preview, and the spout opening is simplified into a<br>blunt capped end, but the whole object reads correctly from normal viewing distance.
Open generated .blend
Six-view review
The six-view teapot uses 20 mesh objects and captures the flattened body and rim<br>stack, but it over-interprets view cues as extra side pads and a vertical front<br>feature. The spout bends into a...