Project Microsite

Off-Road Camping Trailer v2

A custom flat-fronted, tapered-prow box trailer. This record holds the design spec, the drawing workflow, the toolchain, and the project assets.

A steel-framed box camper clad in 4 mm aluminium composite panel (ACP) bonded to the frame, with a flat-fronted tapered nose. The body and nose are one continuous shell — the taper is in plan only; the form is full height throughout. Body 1500 × 1500 × 2000, 2500 overall, on a single off-road axle with externally-mounted wheels and a 550W roof solar panel.

How this evolved

This began as one line — "an architectural drawing of my camping trailer" — and the work visibly changed shape as the goal sharpened. That evolution is the point of keeping versions.

Hand-drawn technical views. First reach was coded orthographic drawings. They kept drifting on precision — pixel geometry typed by hand compounds small errors.
AI text-to-3D. Feeding the written spec to a generator produced a rounded blob (0/10). The lesson: generative models read intent, not millimetres — numbers contribute nothing.
Real photos changed everything. Once you airdropped photos of the actual trailer, image-to-3D reconstructed the true shape, and the messy mesh repaired to watertight.
Right tool per goal. Nano Banana image-to-image gave clean, on-spec pictures in seconds. A picture, an editable mesh, and a dimensioned CAD model are three different jobs — v3 covers the CAD route.

Why start "wrong"? Because the brief was a visual, and the fastest path to a visual is generative AI from text. Watching it fail on geometry is exactly what surfaced the real rule: match the tool to the goal, and feed it real input.

Detailed spec

All dimensions in millimetres.

Main body

Width	1500
Height	1500
Length (rear box)	2000
Overall length	2500
Construction	steel frame + 4mm ACP
Cladding	bonded (glued)

Tapered prow

Nose depth	500
Flat front face	430 wide
Face height	1500 (full)
Position	centred
Taper plane	plan only

Running gear

Tyres	265/70/16
Diameter (approx)	~780
Axle from rear	800
Axle below base	100
Wheel mount	external / outboard

Doors (three)

Rear left leaf	900
Rear right leaf	600
Split	off-centre, full ht
Handles	2 × RV paddle
Front door	RH tapered face

Roof — solar

Panel	550W
Typical size	2278 × 1134 × 35
Mount	flat, centred F–B

Trim (aluminium)

Edge (top/btm/sides)	25 wide
Vertical @ 2000 rear	25 wide
Belt @ 600 base	50 wide

Open item. The 600-from-base belt line needs confirmation — height, extent, and whether it wraps the tapered nose. Front door resolved: third door is set into the right-hand tapered face of the prow.

Drawing & render

For a dimensioned drawing the right tool is parametric CAD — AI image models approximate and will not hold exact millimetres. CAD draws the GA; the renderers handle photoreal.

Fusion 360

Parametric CAD — exact, editable GA drawings

autodesk.com
Recommended — the drawing

FreeCAD

Free / open-source parametric CAD

freecad.org

SketchUp + LayOut

Fast 2D plans / elevations

sketchup.com

Rayon

2D vector design for architects

rayon.design

D5 Render

Real-time ray-traced photoreal

d5render.com

Twinmotion / Unreal

Unreal real-time render, free tier

twinmotion.com

Veras

AI render inside Revit / Rhino / Fusion

evolvelab.io

Midjourney v7

Hero / marketing imagery only

midjourney.com

Prompt → 3D / render

Tools that take the written prompt (or one reference image from it) and output an editable 3D model or a photoreal still. Best-accuracy path is the callout below.

Prompt → editable 3D model (GLB / FBX / OBJ)

Meshy 6

Text & image → 3D, PBR, topology control. Exports GLB/FBX/OBJ.

meshy.ai
Used for the model below

Tripo AI

Fastest (~8s) text/image → 3D

tripo3d.ai

Rodin (3D AI Studio)

Most photoreal from text; rig/remesh

3daistudio.com

Autodesk Wonder 3D

Text & image → 3D in Flow Studio

autodesk.com

Sloyd

Parametric, game-ready 3D

sloyd.ai

Luma Genie

Text → 3D concept geometry

lumalabs.ai

Higgsfield generate_3d

Image → GLB. In your own stack.

higgsfield.ai

Prompt → photoreal still (not dimensional)

Nano Banana (Gemini)

Image-to-image; makes the reference image

gemini.google.com

Flux

High-fidelity text → image

blackforestlabs.ai

Krea

Real-time prompt → image

krea.ai

Leonardo

Image with product presets

leonardo.ai

Best-accuracy workflow. Detailed prompt → 1 reference image (Nano Banana) → image-to-3D (Meshy / Rodin) → export GLB → scale to exact mm in Blender → render in D5 / Twinmotion. One model, fully renderable, correctable.

Clean renders — Nano Banana (image-to-image)

Made from your actual photos: image in, clean image out. A picture, not a 3D model — but fast and on-spec, and the right tool when you just want to see the trailer.

Front render — Front ¾ — tapered prow, roof solar (door placement corrected in prompt v1.0)

Rear render — Rear ¾ — flush barn doors, black RV paddle handles, no external hinges

3D models — routes tested (v2)

Five routes from spec / photos to a renderable 3D model. The lesson: text-to-3D (A, A2) cannot hold the geometry; feeding the real photos (B–D) is what works — and Route B cleaned up best.

Route	Engine	Input	Status
A — text-to-3D	Meshy	spec prompt	Rejected — 0/10
A2 — text-to-3D (hardened)	Meshy	cuboid-first prompt + anti-curve negatives	Succeeded — still weak (text ceiling)
B — multi-image-to-3D	Higgsfield	3 real photos	Best — repaired to watertight
C — multi-image-to-3D	Meshy 6	3 real photos (textured)	Succeeded — archived (9.5MB)
D — single-object (SAM-3)	Higgsfield	front ¾ photo	Failed

Route A — rejected

Meshy text-to-3D from the spec prompt. 0/10 against the brief.

Rejected first-pass mesh — Rejected first pass — archived, not the build geometry.

Why it fails (visual gate). The mesh is a rounded wagon blob: curved tumblehome body instead of a square full-height box; a scooped nose instead of the flat-fronted 430 tapered prow; vague embossed doors instead of the 900 / 600 off-centre barn pair; a flapping roof lid instead of a flat, centred solar panel. Text-to-3D held none of the hard dimensions.

Why A failed. Text-to-3D parses vibe, not numbers — the dimensions contributed nothing; "camper/caravan" pulled the prior toward rounded teardrops; and a feature-stuffed prompt with weak negatives averaged into a blob. A2 fixes the wording (cuboid-first, anti-curve negatives) but the real fix is feeding it photos — routes B–D.

Route B — the winner (Higgsfield multi-image, from your photos)

Reconstructed from the three real photos — the closest result by far. Orbit to inspect.

Repaired — sharp edges

Corrected — watertight, 0 errors

Analysis — your "15 errors": correctable?

Yes — and most cleared automatically. The raw AI export was a soup of unmerged, duplicated faces (it looked solid but wasn't). A repair pass merged vertices and fixed normals and holes; a voxel remesh then forced it fully watertight.

Stage	Broken faces	Boundary edges (holes)	Watertight
Raw AI export	16,943	29,909	No
After repair pass	261	89	No
After watertight remesh	0	0	Yes

Verdict. Fully repairable — all 15 errors clear. But it stays a dense triangle mesh, not parametric; for fabrication-grade dimensions, parametric CAD (v3) is the path. Full-res watertight and Route C exports archived in the vault.

Assets

Trailer photos

The actual trailer — airdropped reference, and the input driving the image-to-3D pass.

Front three-quarter — Front ¾ — tapered prow

Rear three-quarter — Rear ¾ — barn doors

Work log

Session startBrief: architectural drawing of the trailer. Sourced 550W panel dimensions; surveyed AI render + CAD tooling.
GeometryFlat-fronted tapered prow — 430 flat face, full height, centred; body + nose one continuous shell.
Running gear265/70/16 single axle, 800 from rear, 100 below base. Corrected: wheels external / outboard, not under the floor.
Openings & trimThree doors — rear barn pair 900 + 600 off-centre, full height, two RV handles. Edge / vertical / belt trim lines captured.
RoofSolar revised 400W → 550W (2278 × 1134 × 35), centred front-to-back.
ToolingHand-coded drawings drifted on precision. Decision: lock numbers in spec, model via prompt-to-3D, render in D5 / Twinmotion.
3D modelFirst mesh generated via Meshy text-to-3D from the spec prompt.
This siteSpec, toolchain, model and assets logged; shipped to house style.