GPT-6 Astra 3D and Claude Opus 5.5 3D: The Essentials

Last updated: October 25, 2026

GPT-6 Astra 3D and Claude Opus 5.5 3D work in two modes. Build: they turn 1 to 8 reference photos of an object, or just a text prompt, into a real editable 3D asset. Instead of one fused mesh, they reconstruct the object as separately named, selectable parts, each with clean topology and its own PBR material, so a prop can be recolored, swapped, or handed off to Blender or a game engine part by part. Building works best on props and hard-surface subjects such as vehicles, weapons, tools, and furniture. Edit: give them any existing 3D model and a prompt, and they change it: retexture a character, rig it for motion capture, add face shapes for face capture, animate it, unwrap its UVs, or prepare it for a game engine. Every result is delivered with its Blender file, as a GLB by default or in the format you choose. Try them on the GPT-6 Astra 3D model page and the Claude Opus 5.5 3D model page.

Reach for either model whenever the result needs to stay editable, that is, when individual parts must be selected, recolored, or swapped later, and whenever an existing model needs production work that would otherwise take hours in Blender.

Which Model Should I Use?

The two models take the same inputs and share the same seven parameters, so every recipe in this article works on both. They differ in the engine doing the modeling:

  • GPT-6 Astra 3D: the GPT-6 based builder, recently updated and faster than before.

  • Claude Opus 5.5 3D: the new builder powered by Claude Opus 5.5.

Because the inputs are identical, the simplest way to choose is to run the same reference images or the same 3D file through both and compare the results on your own subject. For editing, our 17 side-by-side tests split evenly:

  • Claude Opus 5.5 3D did better on clean unwraps and engine work: the UDIM unwrap, the low-poly rebuild, Unreal prep, fracture, and the courier rig with mocap.

  • GPT-6 Astra 3D did better on precision and motion: the face rig, animation, the mechanical rig, the modular kit, and the head UV.

  • Seven tests were a tie, including the recolor, the retextures and both hair rigs.

The two Side by Side sections below show six building pairs and every editing test.

How to Use the Models

How the Models Work

Everything in this section applies to both models. The building examples below were generated with GPT-6 Astra 3D; the editing examples further down show both models side by side.

To build a new model, feed either model 1 to 8 photos or renders of a single object. The first image sets the viewpoint the result is checked against; any extra images help the model resolve the sides that first photo does not show. Even a single clean reference photo is enough to get a fully separated, editable result.

a vintage brass diving helmet, riveted brass, thick glass viewports

Turntable of the result, built from a single reference photo at Maximum effort. It came back as 33 separate named materials over roughly 151,000 triangles, ready to recolor part by part. Open this asset in Scenario

Feeding Multiple Reference Views

When the object is more complex, add up to 8 photos. Use a front view and a rear (or rear three-quarter) view rather than two side profiles: the model checks the result against the first image, so the extra views should cover ground the first one cannot, not repeat it from another angle.

a compact all-terrain scout drone, matte gray and orange chassis, rugged tires

Turntable built from front and rear reference photos of the same design, capped at a 20,000 triangle game-ready budget with Maximum effort. Open this asset in Scenario

Two independently generated reference photos of the same imagined object can differ slightly in proportions, since each comes from a separate generation. Real photos of one physical object, or renders from a single 3D scene, avoid that drift entirely.

Naming the Subject

When building from photos, the prompt is optional and takes a short name for the object, for example "a tugboat". Leave it blank when the reference photo already shows one clean subject. When building without photos, the prompt is required and describes the whole object, for example "a weathered tugboat with a red hull". Fill it in when the reference has clutter or shows more than one object, so the model knows which one to build. Every building example in this article set it to a short, direct description, for example "a mid-century teak and brass side table" for the side table example below.

Matching Effort and Polycount to the Object

Modeling effort and the polycount cap should track how complex the object actually is, not sit at the same setting for everything. Simple geometric shapes finish in minutes on Standard effort with a low polycount; mechanisms with small moving parts need Maximum effort and a higher polycount cap to hold on to their detail.

"a mid-century teak and brass side table". buildEffort: Standard, faceBudget: 100,000. Open this asset in Scenario

"a cute rounded cartoon toy robot, glossy teal and yellow plastic". buildEffort: Standard, faceBudget: 80,000. Open this asset in Scenario

"a cartoon-style toy food truck with a friendly face". buildEffort: High, faceBudget: 150,000. Open this asset in Scenario

"a steampunk hand-cranked brass flashlight, gear mechanism, leather grip". buildEffort: Maximum, faceBudget: 250,000. Open this asset in Scenario

The side table and toy robot are simple shapes and came back in a few minutes on Standard effort. The flashlight, with its exposed gear mechanism and braided leather grip, needed Maximum effort to keep that detail intact.

Editing an Existing 3D Model

Add any 3D model in the Edit a 3D Model field and describe the change in the prompt. Leave the reference images empty. The model opens the file, makes the change in Blender and hands back the edited model in the format you uploaded, plus the Blender file. Editing works on characters as well as props: in our tests both models recolored and retextured characters, swapped a head, built humanoid, hair and face rigs, animated a rig, unwrapped full-body UDIMs, and prepared vehicles and buildings for game engines.

Write the edit like a brief to a 3D artist. The prompts that worked best in our tests were numbered steps with exact names and measurable targets: bone and object names, triangle budgets, grid sizes, angle limits, UDIM tile numbers, and what must stay unchanged. Every prompt below is the exact text we sent, and both models received the same file and the same prompt.

Match buildEffort to the size of the change, not to the size of the model: Standard was enough for the recolor below, High for retextures and game-ready prep, and Maximum for rigs, animation and the full-body UV unwrap.

Side by Side: Building with Astra and Opus 5.5

To compare the two models fairly, we ran the same reference images, the same subject text, the same effort, and the same polycount cap through both. Below are six of those pairs. The counts under each turntable were read from the delivered GLB files. The Astra assets come from earlier runs of these same inputs.

Vintage brass diving helmet

Inputs: 1 reference photo, Maximum effort, 250,000 triangle cap.

GPT-6 Astra 3D

151,312 triangles, 33 materials, 8.6 MB. Open this asset in Scenario

Claude Opus 5.5 3D

107,252 triangles, 15 materials, 22.7 MB. Open this asset in Scenario

All-terrain scout drone

Inputs: front and rear reference photos, Maximum effort, 20,000 triangle cap.

GPT-6 Astra 3D

19,976 triangles, 11 materials, 1.0 MB. Open this asset in Scenario

Claude Opus 5.5 3D

5,180 triangles, 13 materials, 1.1 MB. Open this asset in Scenario

Mid-century side table

Inputs: 1 reference photo, Standard effort, 100,000 triangle cap.

GPT-6 Astra 3D

12,252 triangles, 6 materials, 2.0 MB. Open this asset in Scenario

Claude Opus 5.5 3D

3,872 triangles, 2 materials, 0.15 MB. Open this asset in Scenario

Cartoon toy robot

Inputs: 1 reference photo, Standard effort, 80,000 triangle cap.

GPT-6 Astra 3D

75,796 triangles, 10 materials, 1.7 MB. Open this asset in Scenario

Claude Opus 5.5 3D

39,272 triangles, 7 materials, 1.0 MB. Open this asset in Scenario

Toy food truck

Inputs: 1 reference photo, High effort, 150,000 triangle cap.

GPT-6 Astra 3D

137,240 triangles, 23 materials, 3.5 MB. Open this asset in Scenario

Claude Opus 5.5 3D

50,917 triangles, 25 materials, 1.2 MB. Open this asset in Scenario

Steampunk hand-cranked flashlight

Inputs: 1 reference photo, Maximum effort, 250,000 triangle cap.

GPT-6 Astra 3D

198,792 triangles, 24 materials, 15.2 MB. Open this asset in Scenario

Claude Opus 5.5 3D

85,732 triangles, 9 materials, 43.1 MB. Open this asset in Scenario

What the six pairs show

  • Opus stayed well under the polycount cap in every pair. It delivered between roughly 4% and 53% of the cap, while Astra used 73% to 100% of it in five of six pairs. Set the cap to the mesh weight you want and expect Opus to come in lighter.

  • Material counts differ by subject. Astra split out more materials on the helmet, side table, toy robot, and flashlight; Opus split out slightly more on the scout drone and food truck. Open both in Blender to see how each model groups the parts.

  • File size does not follow triangle count. The Opus flashlight and helmet came out as larger files than their Astra counterparts, while the simpler Opus pieces were smaller.

  • Judge silhouette and part breakdown on your own subject. Both models take identical inputs, so the fastest test is to run your reference set through each and compare the turntables.

Side by Side: Editing with Opus 5.5 and Astra

We ran 17 edit tests on both models: the same source file, the same prompt and the same effort on each side. Every clip below opens on the source model, then shows Claude Opus 5.5 3D on the left and GPT-6 Astra 3D on the right. The numbers were measured in Blender from the delivered files. Where a clip shows motion capture or swinging hair, we retargeted one standard mocap clip onto each model's rig by bone name and added the same spring settings to every hair chain: the models built the rigs, the motion is ours.

Textures and Materials

Recolor

A color-only change on a character. buildEffort: Standard.

Texture variation only: recolor the jacket from mint green to crimson red, the navy cargo trousers to matte black, and the cobalt backpack to mustard yellow. Keep geometry, pose, proportions and all other details identical.

  • Both models kept the geometry and UVs identical to the source and only repainted the base color map: crimson jacket, black trousers, mustard backpack.

Verdict: tie.

Full Retexture, Three Looks on One Mesh

Three complete texture and PBR rewrites of the same courier: a zombie, a near-future tech courier and a desert scavenger. buildEffort: High.

Texture and material edit only. Keep the geometry, UVs, pose and proportions exactly as they are. Turn this courier into a zombie version of herself: skin on the face, neck and hands becomes pale grey-green with darker bruised veins, sunken dark eye sockets and dried blood around the mouth; the mint jacket becomes filthy and faded, torn-looking, with brown blood stains and mud splatter; the trousers are caked with dried mud up to the knees; buckles, zips and the headlamp turn rusty and dull; the backpack and satchel are scuffed and stained. PBR: cloth, skin and dirt matte and rough, rust rough and non-metallic, only small scratches of bare metal stay slightly metallic. Keep her hair shape and the overall silhouette unchanged.
Texture and material edit only. Keep the geometry, UVs, pose and proportions exactly as they are. Turn this courier into a near-future tech courier: the jacket becomes glossy white and graphite armored fabric with thin panel lines and small printed tech labels; the trousers become a dark carbon-fiber weave; the backpack becomes matte black with a subtle hex pattern; add thin glowing cyan emissive light strips along the jacket seams, the backpack edges and the shoe soles, and make the headlamp lens glow cyan; buckles and metal parts become polished brushed aluminum. PBR: aluminum metallic with low roughness, white panels semi-glossy, carbon weave medium roughness, cyan strips emissive. Keep the face, skin and hair unchanged.
Texture and material edit only. Keep the geometry, UVs, pose and proportions exactly as they are. Turn this courier into a desert scavenger: the jacket becomes sun-bleached sand-colored canvas with patched squares, visible thick stitching and frayed edges; a faded red stencil number 07 on the backpack; straps and belts become worn, cracked brown leather; boots and trousers get a sand-dust gradient from the ground up; tools, buckles and the headlamp become rusted, with scratched bare metal on the edges; the skin gets a light sun tan. PBR: canvas and leather rough, rust rough and non-metallic, scratched bare metal metallic with medium roughness. Keep the hair and face shape unchanged.

  • All six results keep the same 13,204 vertices and identical UVs: only the maps change.

  • Both models rewrote base color, roughness and metallic for every look and added a real cyan emission map for the tech look.

  • Claude Opus 5.5 3D: a fully matte zombie with a blood smear, and a carbon weave on the tech look.

  • GPT-6 Astra 3D: blood dripping from the zombie's mouth, clearer stitched patches on the scavenger, and printed text on the tech jacket that the prompt did not ask for.

Verdict: tie.

Modeling

Head Swap

Cut the head off, retexture it, and reattach it without a seam. buildEffort: High.

Three steps on the head only. 1) Separate the head, including hair, from the body as its own mesh object, cutting cleanly at the neck. 2) Retexture only that head: platinum blond hair and two blue face-paint stripes across each cheek. 3) Reattach the head to the body at the neck so the seam is invisible and the head sits exactly where it was. Keep the body, outfit and pose unchanged, and keep the head as a separate named object in the scene.

  • Both separated the head, gave it platinum hair and blue face-paint stripes, and put it back on the neck with no visible seam.

Verdict: tie.

Fracture for Destruction

Break a crate into Voronoi chunks for a destruction effect. buildEffort: High.

Fracture this wooden crate for a destruction effect in a game. Split it into 20 to 25 Voronoi-style chunks, with more and smaller chunks around the front top corner (the impact point) and larger chunks on the far side. Every chunk must be its own closed, watertight mesh named Crate_Chunk_01, Crate_Chunk_02 and so on, with its origin at the chunk's own center. Keep the original material on the outer faces and assign a new material named Crate_Inside, splintered raw wood, to all newly created inner faces. The chunks must fit together exactly with no gaps and no overlaps, so the assembled crate looks identical to the original. Also keep one untouched copy named Crate_Intact next to it.

  • Claude Opus 5.5 3D: 23 chunks, all watertight, adding up to 99.9% of the crate's volume.

  • GPT-6 Astra 3D: 24 chunks, all watertight, adding up to 100.2% of the crate's volume.

  • Both added the new inside material on the broken faces. A total above 100% means some chunks overlap slightly, which the prompt asked to avoid.

Verdict: Claude Opus 5.5 3D, by a small margin.

Rigging and Animation

Character Rig, Ready for Mocap

A full humanoid rig with fingers, plus secondary bones for hair, bandana and backpack. buildEffort: High.

Rig this character: add a humanoid armature skinned to the mesh with root, hips, spine, chest, neck, head, shoulders, arms, forearms, hands with finger bones, thighs, shins and feet. Add secondary bones for the hair, the neck bandana and the backpack so they can move independently. Keep mesh, pose and textures unchanged.

  • Claude Opus 5.5 3D: 63 bones, 11 of them secondary. GPT-6 Astra 3D: 65 bones, 13 secondary.

  • Both rigs came back skinned, and the same mocap clip drove both without per-rig fixes.

  • Both stretch a little at the shoulder when the arm is raised; the jacket under Astra's armpit stretches more. Rigging from a T-pose base mesh usually cleans this up.

Verdict: Claude Opus 5.5 3D.

Hair Rig: Long Hair and Cape

Mixamo bone names, hair and cape as separate objects, and bone chains down the hair. buildEffort: Maximum.

Production-quality character rig. Be meticulous. 1) Separate the very long waist-length hair (two locks in front of the shoulders, the rest covering the back) into its own mesh object, cut cleanly at the scalp with the hole closed; also separate the short green shoulder cape and the quiver of arrows on the back into their own objects. 2) Build a mocap-ready humanoid skeleton with Mixamo / Unreal Mannequin naming: Hips, Spine, Spine1, Spine2, Neck, Head, Left/Right Shoulder, Arm, ForeArm, Hand, full 3-bone chains for all five fingers, UpLeg, Leg, Foot, ToeBase. Rest pose is the current T-pose, Y-up, root on the floor between the feet. 3) Secondary rig: at least 3 parallel chains of 6+ bones down the hair (left front lock, right front lock, back mass), a 4-bone chain on each side and the center of the cape, one bone for the quiver; hair chains parented to Head, cape and quiver to Spine2. 4) Smooth skin weights with clean deformation at shoulders, elbows, wrists, hips and knees; hair weighted only to hair bones; no stray weights. Keep geometry and textures otherwise unchanged.

  • Claude Opus 5.5 3D: 102 bones. GPT-6 Astra 3D: 107 bones. Both built all 5 hair chains.

  • Because the bones follow Mixamo names, the mocap clip mapped onto both rigs with nothing missing.

Verdict: tie.

Hair Rig: Ponytail and Scarf

The same kind of rig on a second character. buildEffort: Maximum.

Production-quality character rig. Be meticulous. 1) Separate the long ponytail (tied on the crown and hanging down beside the head to the waist) into its own mesh object, cut cleanly at the tie with the hole closed; also separate the long red scarf tail hanging down the chest, both shoulder armor plates, and the katana with its scabbard at the left hip into their own objects. 2) Build a mocap-ready humanoid skeleton with Mixamo / Unreal Mannequin naming: Hips, Spine, Spine1, Spine2, Neck, Head, Left/Right Shoulder, Arm, ForeArm, Hand, full 3-bone chains for all five fingers, UpLeg, Leg, Foot, ToeBase. Rest pose is the current T-pose, Y-up, root on the floor between the feet. 3) Secondary rig: an 8-bone chain down the ponytail parented to Head, a 5-bone chain down the scarf tail parented to Neck, one bone per shoulder plate parented to each Shoulder, and a scabbard bone parented to Hips carrying the katana and scabbard. 4) Smooth skin weights with clean deformation at shoulders, elbows, wrists, hips and knees; ponytail weighted only to ponytail bones; rigid parts weighted 100% to their own bone; no stray weights. Keep geometry and textures otherwise unchanged.

  • Claude Opus 5.5 3D: 68 bones. GPT-6 Astra 3D: 69 bones. Both built an 8-bone ponytail chain and a 5-bone scarf chain.

Verdict: tie.

Face Rig for Face Capture

Facial bones plus all 52 Apple ARKit blendshapes on a head. buildEffort: Maximum.

Production-quality facial rig, ready for facial motion capture. Be meticulous. 1) Separate eyeballs (create them if fused), upper and lower teeth, and a tongue (create one if missing) as their own objects. 2) Skeleton: Neck, Head, Jaw, Tongue (3-bone chain), Left/Right Eye with a look-at target, upper and lower eyelid bones per eye, inner/mid/outer brow bones per side, cheek bones per side, nose bone, lip corner, upper lip and lower lip bones per side. 3) Add the full set of 52 Apple ARKit blendshapes on the head mesh, named exactly per ARKit (eyeBlinkLeft, eyeLookUpLeft, jawOpen, jawLeft, mouthClose, mouthFunnel, mouthPucker, mouthSmileLeft, mouthFrownLeft, browInnerUp, cheekPuff, noseSneerLeft, tongueOut and all the rest) so it works with Live Link Face / iPhone ARKit capture. 4) Clean skin weights around eyes and mouth. Keep the look and textures unchanged.

  • Both delivered all 52 ARKit blendshapes with the exact ARKit names, so an iPhone face capture stream drives them as they are.

  • The skinning is not perfect on either side. The brow bones on Claude Opus 5.5 3D move only 112 vertices; on GPT-6 Astra 3D they move 38,333, and the skin follows every bone.

Verdict: GPT-6 Astra 3D.

Vehicle Rig

Separate wheels, steering, suspension, and a steering wheel wired to the front wheels. buildEffort: Maximum.

Production-quality vehicle rig for a game engine. Be meticulous. 1) Separate each of the four wheels into its own object with its origin at the exact hub center; also separate the steering wheel, the four suspension springs, and the two front steering knuckles. 2) Armature: root, body, suspension_FL/FR/RL/RR (vertical travel only), steer_FL/FR (rotation around the vertical axis only), wheel_FL/FR/RL/RR (spin around the axle only, child of the steer or suspension bone), steering_wheel. 3) Constraints/drivers: rotating steering_wheel drives steer_FL and steer_FR; springs stretch between body and suspension bones. 4) Wheels weighted 100% to their wheel bone, body 100% to body; no stray weights. Unreal Chaos Vehicle friendly naming, Y-up, forward along the vehicle nose. Keep the look and textures unchanged.

  • Claude Opus 5.5 3D: 21 bones; a driver turns the front wheels at 8 to 1 from the steering wheel, with a 30 degree limit.

  • GPT-6 Astra 3D: 29 bones; a driver at 12 to 1 with the angle clamped at 31.5 degrees. It also added 3 demo animations nobody asked for.

  • On both, the suspension compresses on all four corners.

Verdict: tie.

Mechanical Rig: One Slider Opens the Chest

Cut the lid, find the hinges, and drive everything from one Open property. buildEffort: Maximum.

Rig this treasure chest for animation. Cut the domed lid from the base along the seam where they meet, fill the openings so both parts look solid, and separate the front latch hasp with its ring and the two side ring handles into their own objects. Put each object's origin at its real pivot: the lid at its hinge line along the back edge, the latch at its top hinge, each handle at its mounting point. Build an armature with bones root, lid, latch, handle_L and handle_R at those pivots, each part weighted 100 percent to its bone. Limits: lid rotates 0 to 110 degrees around the hinge axis only, latch 0 to 90 degrees, handles swing 0 to 80 degrees. Add a custom property named Open (0 to 1) on the root that drives everything with drivers: first the latch flips up (Open 0 to 0.3), then the lid opens (Open 0.3 to 1.0); handles stay free. Nothing may intersect while opening. Keep the look, geometry and textures otherwise unchanged.

  • On both, Open drives the latch to 90 degrees first, then the lid to 110 degrees.

  • The lid pivot sits 29 mm from the real hinge on Claude Opus 5.5 3D and 4 mm on GPT-6 Astra 3D, which also added an "Open Chest" action.

Verdict: GPT-6 Astra 3D.

Animation: Walk and Idle Loops

Each model animated its own samurai rig: a 24-frame walk loop and a 72-frame breathing idle. buildEffort: Maximum.

Animate this rigged samurai. Keep the mesh, the skeleton, the weights and the rest pose exactly as they are; only add animation. Create two actions on the existing armature: 1) 'Walk_Loop': an in-place walk cycle, 24 frames at 24 fps, a perfect loop where frame 24 matches frame 0; contact poses on frames 0 and 12, passing poses on 6 and 18; planted feet must not slide during contact; arms counter-swing to the legs; subtle hip bob and hip rotation. 2) 'Idle_Loop': a 72-frame breathing idle loop with a slow weight shift from one leg to the other and back, also a perfect loop. In both actions, animate the ponytail and scarf bone chains with follow-through and overlapping action so they lag and settle naturally. Name the actions exactly Walk_Loop and Idle_Loop.

  • Both loops close: the last frame matches the first within 0.000 degrees (Opus) and 0.03 degrees (Astra). The planted foot slides 0.4 cm and 0.6 cm.

  • Claude Opus 5.5 3D swings the ponytail wide but as one block. GPT-6 Astra 3D gives it follow-through, each bone a frame after the one before.

Verdict: GPT-6 Astra 3D.

UV Unwrapping

Head UV Unwrap

Re-unwrap a separated head with the face as one island, using at least 75% of the UV square, then re-bake. buildEffort: High.

Work only on the separated head object. Unwrap its UVs from scratch with seams hidden behind the ears and along the back of the hair, keeping the face as one large undistorted island. Pack all head islands into the standard 0-1 UV square with uniform texel density, minimal stretching, 4-8 px padding, and at least 75% of the square used. Give the head its own material and re-bake its existing texture onto the new layout so it looks identical. Leave the body object, its UVs and its texture untouched.

  • Claude Opus 5.5 3D: 428 islands filling 56% of the square.

  • GPT-6 Astra 3D: 63 islands filling 76%, with the whole face in one island covering 54% of the square.

Verdict: GPT-6 Astra 3D.

Production UV with UDIM Tiles

Unwrap the whole character into five UDIM tiles (head, torso, arms, legs, gear) and re-bake every map. buildEffort: Maximum.

Production UV unwrap of the whole character, from scratch, laid out as UDIM tiles: 1001 head and face with the hair; 1002 torso and jacket; 1003 arms, gloves and hands; 1004 trousers and boots; 1005 backpack, bags, straps and other accessories. Hide seams in natural places: behind the ears and along the hairline, under the arms, on the inner arms and inner legs, along side and back seams of the garments. Keep a uniform texel density inside each tile, with the face at about twice the density of the body. No overlapping islands, about 8 pixels of padding at 2K per tile, islands straightened where it helps. Then re-bake the existing textures (base color, normal, roughness and metallic) onto the new UDIM layout so the character looks identical. Keep the geometry, pose and proportions unchanged.

  • Both put almost every part in the right tile, and the re-baked character looks identical to the source.

  • Claude Opus 5.5 3D: 150 islands, the face at 1.95 times the body's texel density, padding OK, one material per tile.

  • GPT-6 Astra 3D: 853 islands, the face tile at 1.32 times, padding too tight on tile 1005, delivered as true UDIM tiled images.

Verdict: Claude Opus 5.5 3D.

Game-Ready Prep

Low-Poly Rebuild

Rebuild a building for mobile under 5,000 triangles with a single texture atlas. buildEffort: High.

Optimize this building for mobile games: rebuild it as a clean low-poly mesh under 5,000 triangles that keeps the silhouette, roof shape, chimney, door, windows and porch readable. Remove hidden and interior faces. Bake the fine detail from the original high-poly onto a tangent-space normal map. Use a single material with one 1024x1024 texture atlas (base color, normal, ORM packed) in the same art style and colors as the original.

  • Claude Opus 5.5 3D: 342 triangles, atlas 66% used.

  • GPT-6 Astra 3D: 1,370 triangles, atlas 39% used with empty tiles.

Verdict: Claude Opus 5.5 3D.

Modular Kit

Split the same building into reusable kit pieces on a 1 m grid, laid out 1 m apart. buildEffort: High.

Turn this building into a modular environment kit for level design. Split it into reusable separate objects: straight wall, wall with window, wall with door, outer corner, roof section, roof ridge/edge, chimney, porch floor, porch post, barrel, hanging sign. Make wall pieces exactly 2 m wide on a 1 m grid so they tile seamlessly; set each piece's origin at its bottom-left-back corner on the grid. Lay the pieces out in a neat row with 1 m spacing, named kit_wall_straight, kit_wall_window, etc., and keep one fully assembled copy of the tavern built from those pieces next to them. Same textures and style.

  • Claude Opus 5.5 3D: gaps between pieces from 1.0 to 2.4 m.

  • GPT-6 Astra 3D: exactly 1.0 m, as asked.

Verdict: GPT-6 Astra 3D.

LOD Chain

LOD0 to LOD3 at about 100, 50, 15 and 5% of the triangles. buildEffort: High.

Create a game-ready LOD chain for this vehicle as separate objects in the same scene: buggy_LOD0 (the original), buggy_LOD1 at about 50% of the triangles, buggy_LOD2 at about 15%, buggy_LOD3 at about 5%. Every LOD shares the same UV layout and material, keeps the silhouette and wheel shapes as long as possible, and keeps the wheels as separate objects down to LOD2. Place them side by side, 1 m apart.

  • Both hit 100 / 50 / 15 / 5% with the same UVs and material, and kept the wheels separate down to LOD2.

Verdict: tie.

Unreal Engine Prep

Separate parts with SM_ names, real scale and axes, UCX_ collision hulls, a lightmap UV and sockets. buildEffort: High.

Prepare this buggy as a game-ready asset for Unreal Engine 5. 1) Separate the body and the four wheels into their own meshes, each wheel with its origin at the exact hub center. 2) Real-world scale (about 3 m long), Z up, front of the vehicle facing +X, the body pivot on the ground plane at the center between the four wheels; apply all transforms. 3) Name meshes with the SM_ prefix: SM_Buggy_Body, SM_Buggy_Wheel_FL, SM_Buggy_Wheel_FR, SM_Buggy_Wheel_RL, SM_Buggy_Wheel_RR. 4) Add simplified convex collision hulls following Unreal naming: up to 8 hulls for the body named UCX_SM_Buggy_Body_00, UCX_SM_Buggy_Body_01 and so on, and one hull per wheel named UCX_SM_Buggy_Wheel_FL_00 etc. 5) Add a second UV channel named UVLightmap on every mesh: no overlapping islands, with padding, all inside 0 to 1. 6) Add sockets as empties named SOCKET_Wheel_FL, SOCKET_Wheel_FR, SOCKET_Wheel_RL, SOCKET_Wheel_RR at the hubs, SOCKET_Driver_Seat on the seat, and SOCKET_Exhaust at the exhaust or rear. Keep the materials and textures unchanged.

  • Both: 5 SM_ meshes, 3.00 m long, Z up, front on +X, 6 sockets. Claude Opus 5.5 3D built 9 collision hulls, GPT-6 Astra 3D 12.

  • GPT-6 Astra 3D swapped left and right: its SM_Buggy_Wheel_FL is really the front right wheel.

  • Lightmap UVs have no overlaps on either, but use 11% of the square on Opus and 41% on Astra.

Verdict: Claude Opus 5.5 3D.

More Edits from Claude Opus 5.5 3D

Two more edits, both run on Claude Opus 5.5 3D only. They show edits that add or reshape geometry rather than change textures or rigs.

Adding New Parts

A robot built with another model, edited to add booster nacelles, a turbine backpack, a tractor-beam emitter and a planter. buildEffort: High.

Edit this existing maintenance hover robot for a premium solarpunk exploration game. Preserve its friendly compact ivory and orange body, black face and two cyan eyes, UVs and existing textures. Add two substantial articulated booster nacelles on the left and right rear sides with dark graphite cylindrical exhaust bells pointing downward, copper collars, cyan emissive inner cores and segmented ivory cowling. Add a compact rear turbine backpack, a small front-facing tractor-beam emitter beneath the face, fine orange maintenance markings, and one small green plant sprout in a protected glass-like dorsal planter. Keep a clean readable silhouette. New parts must be named and use 5-8 shared PBR materials. Real solid low-poly geometry, no floating text or ground plane. Keep the mesh lightweight; use moderate radial segments for cylinders. Preserve original proportions and orientation. Deliver the complete edited robot.

The source had 12,206 faces and one material; the edit came back at 20,510 faces with 8 materials.

Transforming the Object

A short prompt can reshape the whole object while keeping its style. buildEffort: High.

transform this 3D model to a car instead, in the same style

The toy biplane became a toy car with the same yellow and red palette and wheels.

What the 17 Edit Tests Show

  • Claude Opus 5.5 3D did better on 5: the courier rig with mocap, the low-poly rebuild, the UDIM unwrap, Unreal prep and fracture.

  • GPT-6 Astra 3D did better on 5: the head UV, the face rig, the modular kit, animation and the mechanical rig.

  • 7 were a tie: recolor, retexture, head swap, both hair rigs, the vehicle rig and the LOD chain.

  • Both hand back rigs ready for capture: body skeletons with Mixamo bone names such as Hips, Spine2, LeftArm and RightFoot, and faces with all 52 ARKit shapes such as jawOpen, eyeBlinkLeft and mouthSmileLeft.

Parameters

Seven parameters control both models: what to build from or what to edit, what to build or change, how hard to work at it, how many triangles to deliver when building, and the file format and face type of the result.

images

Optional. When building, add 1 to 8 photos or renders of a single object, or leave it empty to build from the prompt alone. When editing, add images only to guide the change, such as a material sample, a logo or a marked-up view of the model. The first image sets the viewpoint the result is checked against; the rest help resolve the sides that first photo does not show. The scout drone example uses two views (front and rear); the diving helmet example reconstructs a full object from just one.

sceneFile

Edit a 3D Model. Optional: add any 3D model to edit it instead of building a new one, and it comes back in the same format. Every example in the editing section used it, from a single-mesh crate in the fracture example to a 380,000-vertex head in the face rig example.

prompt

Required when editing: describe the change, like "open the lid". The more exact the brief, the better: the chest rig example named every part, hinge and angle limit. When building without photos, it is required: describe the object, for example "a weathered tugboat with a red hull". When building from photos, it is optional: name the object in a few words, for example "a tugboat". Leave it blank when the reference photo already shows a single clean subject; fill it in when the reference has clutter or shows more than one object, so the model knows which one to build. The side table example used "a mid-century teak and brass side table".

buildEffort

Standard, High, or Maximum, default High. When building, match it to how complex the object actually is instead of leaving it maxed out for everything. When editing, match it to the change, not the model: the recolor used Standard, and the face rig used Maximum. The side table and toy robot examples used Standard and finished in a few minutes; the steampunk flashlight, with its exposed gear mechanism, needed Maximum to hold on to the small mechanical detail.

faceBudget

1,000 to 2,000,000 triangles, default 250,000. Used only when building; when editing, put any triangle target in the prompt instead, as the low-poly example did. This is a hard cap applied at export, not a suggestion. The scout drone example was capped at 20,000 for a game-ready budget and still kept its wheels and sensor turret readable; the flashlight example used the full 250,000 to preserve its gear teeth and rivets.

outputFormat

Auto, GLB, FBX, OBJ, USDZ, STL, PLY or BLEND, default Auto. Every result also comes with its Blender file. Auto delivers a GLB when building and keeps the edited model's own format when editing. With Quads topology, the result is an FBX unless the model is already FBX, OBJ, PLY or USDZ.

topology

Auto, Triangles or Quads, default Auto. The faces of the delivered model are checked before it is returned. Auto keeps an edited model's faces and lets the model choose for a new one.

Use Cases

  • Game props: build weapons, vehicles, and set dressing as editable meshes with parts already separated for material variants.

  • Product visualization: turn product photography into a 3D asset for interactive viewers or configurators.

  • Film and VFX previz: convert concept photos or physical maquettes of props into a scene-ready starting mesh.

  • E-commerce: give shoppers a rotatable 3D view of a real product straight from its catalog photos.

  • Education and museums: digitize physical objects or replicas into inspectable, part-labeled 3D models.

  • Character pipelines: rig existing characters with Mixamo bone names for motion capture, add ARKit face shapes for iPhone face capture, and generate walk and idle loops.

  • Game-ready conversion: turn a detailed model into a low-poly mesh, a modular kit, a LOD chain, or an Unreal-ready asset with collision and sockets.

  • Skins and variants: retexture one character or prop into several looks on the same mesh and UVs.

  • Tabletop and print-and-play: turn concept art or physical minis into printable terrain and prop kits.

Tips for Better Results

  1. Use front-and-back reference views, not two side profiles, for multi-image input. The model checks the result against the first image and needs the others to cover what that first view cannot show.

  2. Fill in the Subject field whenever the reference photo has clutter or more than one object, even though the field is optional.

  3. Match buildEffort to the object's real complexity. Simple geometric shapes finish in minutes on Standard; mechanisms with small moving parts benefit from Maximum.

  4. Set faceBudget to your pipeline's real target up front. It is a hard export cap, so a game-ready asset and a hero-detail asset should use different values, not the same default.

  5. Shoot or generate reference photos with clean, even lighting and an uncluttered background. The model separates the object from its surroundings more reliably that way.

  6. Expect strong per-part material separation, often a couple dozen named materials, even from a single reference photo. That separation is what makes the result editable afterward.

  7. Write edits as numbered steps with exact names and numbers. Bone names, part names, triangle budgets, grid sizes, UDIM tiles and angle limits gave the most predictable edits in our tests.

  8. Say what must not change. "Keep geometry, pose and UVs identical" kept every texture edit on the original mesh, byte for byte.

  9. Ask for Mixamo or ARKit names when the rig is headed for capture. Both models followed the naming exactly, so mocap and face capture data mapped on with no renaming.

  10. Rig from a T-pose base mesh. Both rigs stretched a little at the shoulder on a relaxed-pose character; a T-pose base usually cleans this up.

  11. Use Maximum effort for fine structures like braided leather straps, thin crank handles, or rivets; they hold up well at higher settings.

Known Limitations

  • Independently generated reference photos of an imagined object can drift slightly in proportion between views, since each comes from a separate generation. Real photos of one physical object, or renders from a single 3D scene, avoid this.

  • Building from images is made for props and hard-surface subjects. To work on a character, build or bring the mesh from a character model, then use edit mode to retexture, rig, animate or unwrap it.

  • Skinning on edited rigs is a strong start, not a final pass. Shoulders stretched a little on both character rigs, and the Opus face rig left the brow skin almost still. Check deformation in Blender before shipping.

  • Edits can add things nobody asked for. In our tests GPT-6 Astra 3D printed text on a retextured jacket and added demo animations to a vehicle rig. Say "nothing else" in the prompt and review the file.

  • Check names, sides and packing before export. GPT-6 Astra 3D swapped the left and right wheel names in the Unreal test and packed one UDIM tile too tightly; Claude Opus 5.5 3D used only 11% of the lightmap UV square.

  • Surfaces never shown in any reference photo are inferred, not observed. Treat a fully hidden underside or interior as a best guess, not a measurement.

  • Maximum effort with a high faceBudget takes real time, often tens of minutes per asset in testing. Plan batch runs accordingly, and start with Standard or High effort when you only need a quick look.