Production test report

AI-accelerated elevation renderings:
what we tested, and what works.

Four production workflows were tested on real builder homes — from pure AI-from-drawings to AI applying materials over a fully detailed render. This report documents each workflow, where its output diverges from the drawings, and the workflow recommended for production.
Read the summary ↓ Jump to recommendation
Summary of findings

Four workflows tested. One holds up in production.

Three of the four workflows let AI touch the house itself. In each, AI re-invents geometry or materials, leaving 7–40% of every elevation wrong in ways no prompt clears — small but unpredictable hand-work on every home. The fourth keeps AI off the house entirely: build the home in human-made 3D, render on a bare base, and let AI build only the environment and sky, then composite and upscale to 6K. It is the only workflow that keeps geometry, materials and camera fully controlled.

On render style — this test was not about choosing a final look. Its only goal was to find the approach that keeps four components consistent and correct: geometry, materials, camera & scale, and landscaping / environment / sky. The visual style of the final render is open — any client preference can be applied at the styling stage without affecting how well the method holds those four.

40
House types
6K
Delivery res.
4
Workflows tested
7–40%
Defects when AI touches the house
10–15
Facades / week (target)
builder home render
The test method

Three things under our control on every home.

Every workflow is measured against the same three assets: the builder’s drawing and material spec (ground truth), a human-made 3D model authored to that drawing, and the controlled render it produces. (Workflows were tested on more than one real builder home; the example below is a Fieldgate Caledon model.)

InputCAD elevation + material spec
A · Drawing + material spec
Builder’s CAD elevation set and material / colour schedule — the source of truth for every opening, trim, dimension and material.
Ground truth
Authoredhuman-made 3D model
B · Human-made 3D model
Full geometry built to the drawing — walls, openings, roof, trim, all exact. Untextured clay pass shown.
100% our control
Renderedcontrolled render
C · Controlled render
Materials & mapping from the project library, our camera and light. Pixel-accurate to the drawing.
100% our control
How to read each workflow

Each workflow below tries to replace part of B–C with AI. We show the workflow sequence, the AI outputs it returns, a scorecard against five criteria, and the technology used. The gap between the output and the drawing is the hand-work each workflow leaves behind.

Workflow 1 · pure AI from the drawings

Pure AI from drawings: the output varies on every run.

The elevation, a visual reference and the material spec all go to the AI model — with a detailed prompt spelling out where each material belongs (not just a spec screenshot left for the AI to interpret). Even so, the output drifts randomly in both geometry and materials and varies on every run; accuracy tops out around 70%.

The workflow — three inputs
Drawingsdrawings
Drawings
Builder’s elevation.
+
Visual refvisual reference
Visual reference
For sky, light & environment.
+
Material specmaterial spec
Material spec
+ detailed prompt on placement.
➞
AI+ detailed prompt
AI outputs — same input, random drift every run
Pass 1pass 1
Geometry + materials off
Massing, materials & angle all drift from the drawing.
AI
Pass 2pass 2
Geometry + materials off
Different geometry, materials & camera again.
AI
Pass 3pass 3
Geometry + materials off
Another set of geometry & material shifts.
AI
Pass 4pass 4
Geometry + materials off
Geometry, materials & angle drift once more.
AI
Workflow 1 scorecard
Draft estimates — confirm / adjust with the team
Geometric accuracy
35–65%
Material accuracy
30–60%
Consistent camera & scene scale
20–45%
Foreground landscaping
Good
Background / sky
Excellent
Conclusion — nothing constrains the structure, so geometry, materials and camera drift randomly run to run — around 70% accuracy at best. Fastest workflow, but it cannot anchor a consistent lineup.
Technology

text-to-image / image-to-image with a visual reference and a material-placement prompt. Drawings carry title-block lines, dimension text and downspouts the model reads as geometry; with no structural lock, geometry and materials are re-drawn randomly each run — a detailed prompt narrows it but cannot hold it.

Workflow 2 · simplified blockout + projected drawing

Simplified blockout, drawing projected, then AI for the skin.

A light blockout (no window or door openings) with the elevation projected onto the front facade, plus the material spec (reinforced in the prompt) and a visual reference render — all handed to AI for materials and detail. The camera stays fixed, but geometry and materials drift.

The workflow — four inputs
Blockoutblockout
Simplified model
Light blockout, no openings.
Our team
+
Projectionprojected render
Projected render
Drawing mapped on, fixed camera.
Our team
+
Material specmaterial spec
Material spec
+ prompt reinforcement.
+
Visual refvisual reference
Visual reference
For look & environment.
➞
AImaterials + detail
AI outputs — same base, divergent results
Output 1output 1
Output 1
Geometry drift — a wall appears beside the left corner 2nd-floor window, which the drawing carries to the corner. Many materials misread.
AI
Output 2output 2
Output 2
Windows roughly right, but an invented gable & sloped roof appears and the roofline over the garages is wrong. Materials still off.
AI
Output 3output 3
Output 3
Materials & trim change again.
AI
Workflow 2 scorecard
Draft estimates — confirm / adjust with the team
Geometric accuracy
70–88%
Material accuracy
55–78%
Consistent camera & scene scale
90–100%
Foreground landscaping
Good
Background / sky
Excellent
Conclusion — the blockout fixes the camera, but with no openings the AI invents every window, door and trim. Geometry and materials still need a 7–15% hand-fix on each home.
Technology

3ds Max blockout + projected elevation; depth/edge guides (ControlNet depth + MLSD) hold the massing, but fine trim leaks and material UV scale is not stable home-to-home.

Workflow 3 · fully detailed model + AI materials

Fully detailed model, AI applies the materials.

The home is built as a full detailed, honestly-textured 3D model; AI is given the material spec (reinforced in the prompt) and a visual reference for planting, and applies the finish. Geometry and camera hold — but the materials skew and drift pass to pass.

The workflow — three inputs
Detailed modeldetailed model
Detailed model
Full honest-material model, clay pass.
Our team
+
Material specmaterial spec
Material spec
+ prompt reinforcement.
+
Visual refvisual reference
Visual reference
For landscaping & planting.
➞
AIapply materials
AI outputs — 3 material passes on the same model
Pass 1material 1
Material pass 1
Brick + grey board-batten & stone.
AI
Pass 2material 2
Material pass 2
White gable; brick & trim shift.
AI
Pass 3material 3
Material pass 3
Grey stucco; banding & lintels change.
AI
Across all three passes — geometric drift is minor: proportions of individual architectural elements shift slightly and are barely noticeable to the eye. But material drift, and the degree to which materials diverge from the spec, stays significant in every output.
Workflow 3 scorecard
Draft estimates — confirm / adjust with the team
Geometric accuracy
85–93%
Material accuracy
65–85%
Consistent camera & scene scale
90–100%
Foreground landscaping
Good
Background / sky
Excellent
Conclusion — geometry holds well (85–93%), but AI material transfer is not UV-aware, so materials and fine detail still need a 7–15% hand-fix per home.
Technology

AI relight / material transfer over a finished render; not UV-aware, so seamless textures mis-handle openings, trim and reflective glazing — correctness can’t be guaranteed.

Workflow 4 · human-made 3D + AI environment — recommended

Human-made 3D for the house, AI for the environment only.

The house is a human-made render we control fully; AI builds only the surrounding environment and sky, then we composite and upscale to 6K. The house and camera are identical across every pass — only the world changes.

The workflow
Drawingdrawing
Drawing
Source of truth.
Our team
➞
Modelmodel
Human-made model
Full geometry & detail.
Our team
➞
Bare renderbare render
Bare render
Black background, no planting.
Our team
➞
AIenvironment only
AI outputs — house & camera identical, only the world changes
+ AI env 1env 1
Environment 1
House untouched.
AI
+ AI env 2env 2
Environment 2
Different planting & sky.
AI
Final · 6Kfinal example 1
Final example 1
Composited & upscaled to 6K.
AI env
Final · 6Kfinal example 2
Final example 2
Different environment, same house.
AI env
↓  Download example 1 · full 6K (ZIP) ↓  Download example 2 · full 6K (ZIP)
Workflow 4 scorecard · recommended
Draft estimates — the house is 100% human-made geometry; AI only touches the world
Geometric accuracy
100%
Material accuracy
100%
Consistent camera & scene scale
100%
Foreground landscaping
Excellent
Background / sky
Excellent
Conclusion — the house is exact and revisions are deterministic (edit the model, re-render). After setup the line targets 10–15 finished hi-res facades per week. This is the recommended production workflow.
Before / after — bare render to final, drag to compare
final render with AI environment bare render, black background
Bare render
Final · AI environment
↔
Technology

3ds Max + Corona, locked real-scale material library, one fixed camera for the set. AI environment relight over a clean beauty pass, composited in Photoshop and reconciled to the source. Final 2–4× AI upscale to native 6K — no generative model outputs 6K directly.

Derivative outputs

Human-made models also produce outputs AI renders cannot.

Because every house is a real 3D model, the same assets drop into any shared scene. Streetscapes and aerials need several homes together under one light, one scale and one perspective — straightforward with real models, and not possible from AI facade renders, which are isolated flat images.

Streetscapestreetscape
Multiple distinct models in one scene
One light, one perspective, consistent scale — only possible from human-made 3D.
Aerial / blockaerial
Aerials & bird’s-eye
Same models, a new camera — no re-work.
Derivatives reuse the models

Streetviews and aerials reuse the exact same models — no re-modelling, no AI guesswork. The same library carries into marketing, signage and future phases.

Interiors, in parallel

Interiors are handled by dedicated interior artists, running alongside the facade line — one team, one timeline, one point of accountability.

Risk & recommendation

Where AI belongs in the workflow — and where it does not.

✓
AI as finisher of the environment
  • The house is 100% human-made geometry — zero defects in the controlled layer.
  • Revisions are deterministic: edit the model, re-render, done.
  • Client changes land in a controlled scene, not a generative black box.
  • Throughput is estimable — 10–15 facades/week after setup.
✗
AI as author of geometry / materials
  • 7–40% of every elevation is wrong — and which share moves each run.
  • Each defect is hand-fixed — it does not scale to 40 homes.
  • Client revisions happen in an uncontrolled AI scene — the highest risk.
  • Throughput is unknowable — the rescue work cannot be quoted.
Recommendation
Direct resources to full human-made modelling & texturing — not to correcting unpredictable AI defects.
Model the houses

Build full-detail human-made 3D houses across the lineup; add modellers (outsource if needed).

AI for the environment

Let AI do only environment, planting & sky — the work it does fast and safely.

Deliver at 6K

Composite, reconcile and upscale to native 6K, client-ready, fully under control.

×