G—P / FMX 430 / Semester / Week 05

Week 05 · Phase 1 — World Building

Architectural visualization pipeline Project 2 launch

Until now you've built worlds from scratch. Archviz is the opposite discipline: someone else's geometry arrives — CAD, BIM, a Twinmotion scene — and your job is to move it into Unreal without mangling it, then photograph it like a client is paying. Project 2 launches today, and it's the most employable week of the semester.

Session Fri Oct 02 · 10:00 am – 1:50 pm · CCB-134 Bring your laptop with Twinmotion installed and working
By the end of this session you can

2:00 · Brief Other people's geometry

The archviz pipeline is a translation problem. Architects work in CAD and BIM tools that care about tolerances and revisions; game engines care about triangles, UVs, and frame time. Datasmith is Epic's translation layer — it carries geometry, materials, hierarchy, and metadata from those worlds into Unreal, and Twinmotion sits on the same foundation as the fast, designer-friendly front end. Project 2 (see the Projects tab) lives in this pipeline for the next five weeks.

What Datasmith preserves — and what it can't

Expect the scene hierarchy, instancing, and material assignments to arrive largely intact; expect to do surgery on everything a CAD tool never had to think about. The reliable wound list: object names like Extrusion_2847, no collision anywhere, lightmap UVs missing or overlapping, pivots at the world origin instead of on the object, and glass that imports as a grey slab. None of these are bugs — they're the difference between a drawing and a game engine. Budget for cleanup; the brief's "pipeline notes" deliverable is exactly this ledger.

Archviz is photography

What a client buys is not the model — it's the image. Which makes this week secretly a photography class: eye-level cameras (not drone-hovering god views), verticals kept vertical, wide but honest lenses in the 18–35 mm range, and depth of field used to direct the eye rather than to show off. In UE that language lives in the Cine Camera Actor, with real-world focal lengths and apertures, and in 5.8's rendering path for stills and flythroughs with accumulated, film-quality depth of field.

Scan versus model

The industry's live question, and this week's discussion: capture is getting so good that sometimes the right move is to not model at all. That's where Gaussian splatting comes in.

State of the Art · Scan vs. model — Gaussian splatting

3D Gaussian Splatting (Kerbl et al., SIGGRAPH 2023 — this course's landmark paper) renders captured reality at 100+ fps and is now the fastest-growing capture method, entering UE through plugins like Volinga and NanoGS (no first-party Epic support yet). LiDAR and photogrammetry still own measurement. Discussion: for an archviz client, when do you scan, when do you model, and when do you combine?

Kerbl et al. project page · Splats in UE, 2026

2:45 · Guided lab Round trip: Twinmotion → Unreal → camera

Everyone runs the same pipeline end to end today with a provided architectural model, so the workflow is proven before you commit to your own Project 2 subject.

  1. Dress the scene in Twinmotion

    Open the provided model in Twinmotion. Spend twenty minutes doing what Twinmotion is best at: drag-and-drop materials, vegetation, a time of day. Note how fast iteration feels — that speed is why the pipeline starts here.

  2. Export for Unreal

    Export the scene from Twinmotion as a Datasmith file (via its export options for Unreal / Datasmith). Note what the export dialog offers to include — that list is exactly what will and won't survive.

  3. Enable Datasmith in UE

    In a fresh UE project (FMX430_Archviz, under Git before anything is imported): Edit → Plugins, enable the Datasmith importer plugins, restart if prompted.

  4. Import and inspect the damage

    Import the Datasmith file from the main toolbar's Datasmith import. Then inspect like a surveyor: open the content browser folders it created, check names, select meshes and look for collision and lightmap UVs in the static mesh editor. Write down the top five problems — this list drives the AI lab.

  5. Fix a few by hand first

    Rename a handful of assets properly, generate collision on one mesh in the static mesh editor, fix one material assignment. You need the manual feel for each fix before you delegate it at scale.

  6. Shoot a hero still

    Place a Cine Camera Actor at standing eye height. Choose a focal length deliberately (start near 24 mm), level the horizon, compose with foreground. Set aperture and focus distance in the Details panel so depth of field serves the composition. Render a high-resolution still.

  7. Block a flythrough

    Rough out a 20–30 second camera move through the space (Sequencer basics on the projector — full cinematics come later in the course). Slow is professional; fast is a video game trailer.

  8. Commit everything

    Repo committed and pushed, including your import problem list — it's the seed of Project 2's pipeline notes.

4:00 · AI lab Batch cleanup, on the record

Import cleanup is the perfect agent task: repetitive, rule-based, verifiable — and dangerous at scale, because a bad rule applied two hundred times is two hundred new problems. Today's theme is the ledger: what it fixed, what it broke.

Prompt 1 · The audit
Audit the assets this Datasmith import created. List: meshes with auto-generated or meaningless names, meshes without collision, meshes with missing or overlapping lightmap UVs, and anything else that would embarrass a professional archviz project. Don't change anything yet — report first.

Compare its audit against the problem list you wrote by hand in the guided lab. What did it catch that you missed? What did you catch that it missed? Both numbers go in the log.

Prompt 2 · The batch fix
Fix the naming first: rename the imported static meshes to the SM_ convention with descriptive names based on what each mesh is. Then generate simple collision for the architectural meshes. Work in batches of ten and report after each batch.

Batches are the safety mechanism — review each one before allowing the next. Spot-check renames against the actual geometry (did it name a column SM_Column or guess wrong?) and open two meshes to confirm collision actually exists.

Prompt 3 · The damage report
Diff your changes against the original import: list everything you modified, anything you could not fix, and anything you may have made worse. Be specific about the uncertain ones.

Then run git diff --stat yourself and compare. An agent's self-report and the repo's actual diff never quite match — the gap between them is this week's most honest data.

Before you call it done

5:15 · Crit / share Hero stills

Everyone puts their one hero still on the projector. We read them as photographs: where does the eye go, are the verticals vertical, what is the depth of field for? Declare your Project 2 subject — an unbuilt or reimagined space — before you leave. And remember the schedule: Project 1 is due at crit next week, so this weekend serves two masters.

Homework — due before Week 6

TaskDeliverable
Finish Project 1: final materials, lighting, and a 60-second flythrough or walkable build — it's due at crit next week.Complete Project 1 in the repo: build, reference sheet, and AI development log.
Source the model for your Project 2 subject and run it through the pipeline once.Datasmith import in your archviz repo, plus your initial import problem list.
Read the Gaussian splatting project page (abstract and results at minimum).Three sentences in your log: a case where you'd scan instead of model, and why.

Resources