Skip to content
newblender

11 — Nashville International Airport Grand Lobby: how an agent could make it
Original Gentilhomme (Montreal), 2023, immersive motion graphics for two large-format LED walls · blender.org user story
Agent-readiness today Partial: the Geometry Nodes (GN) motion-design core, Cycles multipass output and scripted rendering work through the live bridge today. The 40-minute volume at 11,200 px wide needs a headless farm (L-011), and the design language (neon, distilleries, country music) is owner taste.
Difficulty 4 (one capsule is a week; the full 40+ minutes is a studio-quarter project, dominated by render compute and art direction)
First slice A 10-second seamless “Broadway neon grid” loop, built as one GN tree, framed for the 11,200 × 2,160 wall and rendered at quarter resolution (2,800 × 540) to 32-bit multilayer EXR, with a check script that proves the loop closes and nothing clips.

1. What the original actually is

Section titled “1. What the original actually is”

Facts from the blender.org user story (fetched 2026-10-07):

  • Display: two LED walls in the Grand Lobby, each 21 m × 4.5 m. Standard canvas 7,360 × 2,160 px per wall; the maximum canvas is 11,200 × 2,160 px (an aspect of about 5.2 : 1).
  • Volume: over 40 minutes of content in 13 segments.
  • CGI share: 6 segments (“capsules”) are fully CGI; themes include distilleries, country music and Broadway-style neon.
  • Two content families: “traditional CGI” (rigging, animation, simulation) and motion design with slow ambient animation.
  • Output: multipass EXR in 16- or 32-bit, composited in After Effects.

Deliverables, restated as an agent would track them: 13 timed segments × 2 walls; per segment a frame sequence at wall resolution, per-pass EXRs, a composited master, and loop points where the segment cycles in the lobby.

2. How the humans made it

Section titled “2. How the humans made it”

Stages exercised: S03 Shading, S04 Geometry Nodes, S09 Lighting & rendering, S10 Compositing.

Fact (user story):

  • Cycles was the render engine. GN was used “for creating diverse visual elements”: grid distributions and texture-driven displacement, per the index entry.
  • Neon tube assets were built in SideFX Houdini from Adobe Illustrator designs. Other assets came from Maya and Substance Designer. Smoke came from EmberGen.
  • A 20-machine render farm ran custom scripts (“GH render” and “GH multipass”) written by in-house creative coders. Denoising ran as part of the render-node process.
  • Renders of a capsule at final resolution took “full nights or even several days”.
  • Team: CG supervisor Arnaud Mellinger, motion designer Maxime Roux, plus modelling, coding and Houdini specialists.

Interpretation: the studio already worked in an agent-shaped way for the parts that scale: procedural layouts in GN, scripted farm submission, scripted pass handling. The hand-made parts were design (Illustrator), hero assets (Houdini/Maya) and the look in After Effects.

3. How I would make it: the agent-native plan

Section titled “3. How I would make it: the agent-native plan”
Stage Agent approach Inputs (free sources) Tooling
S04 Geometry Nodes One GN tree per capsule, the primary artefact (L-006): Grid → Instance on Points → per-instance phase from Random Value; Set Position driven by a Noise Texture sampled at (x, y, t) where t loops on a circle so the animation closes Scene Time node; no external data tools/live/bl.py builds the tree with bpy.data.node_groups.new(..., "GeometryNodeTree")
S01 Modeling (neon) Neon letters as curves → Curve to Mesh with a circle profile, all in GN; letter shapes from a font via Text object → Curve, or SVG import Free OFL fonts (Google Fonts), CC0 SVG GN String to Curves node
S03 Shading Emission shader with strength from a named attribute (flicker) written by GN; glass tubes as thin refractive shells None Shader Attribute node
S08 FX (smoke) Not EmberGen: Blender Mantaflow smoke baked headless, or a cheap volumetric noise shader for ambient haze None Headless bake
S09 Rendering Cycles, fixed seed, multilayer EXR, passes: Combined, Emission, Denoising Albedo/Normal, Cryptomatte Poly Haven HDRI only if a reflection environment is wanted (CC0) Headless blender -b per frame range
S10 Compositing Blender compositor node tree written as code: glare (fog glow) on Emission, grade, output to the delivery codec None Compositor via Python
Farm Split by frame range across machines; split very wide frames by render border and stitch Local GPUs Shell + blender -b -P

Build order:

  1. Write the checks first (§5): resolution, frame count, loop closure, clipping, determinism.
  2. Set the canvas: 11,200 × 2,160 at 25% for look-dev, metric scene scale matching the 21 m wall.
  3. Build the GN grid-and-displace tree on an empty object; declare every parameter as a group input (spacing, amplitude, speed, palette index).
  4. Write the emission material reading the GN attribute.
  5. Camera: orthographic, sized to the wall aspect, so pixels map 1:1 to the LED grid.
  6. Render 3 test frames (0, mid, last) to EXR; run the checks; snap the viewport for the owner (tools/live/snap.py).
  7. Owner reviews look; parameter edits only, no topology edits.
  8. Full-resolution render by frame range headless; stitch borders; composite.

4. What works today vs. what needs newblender

Section titled “4. What works today vs. what needs newblender”
  • Works today with the live bridge:
    • Building GN trees node by node from Python and wiring group inputs. The chair demo already showed declared modifiers built and checked live (~/newblender-data/demos/001-chair/report.md, 12/12 checks).
    • Setting Cycles samples, seed, resolution, passes and multilayer EXR from Python.
    • Reading a rendered EXR back and measuring pixels (bpy.data.images.load, pixels.foreach_get).
    • Screenshots of the owner’s window for the review loop (tools/live/snap.py).
  • Painful today:
    • Wiring large GN trees through nodes.new / links.new is verbose and position-less. The tree is the artefact (L-006), but there is no first-class text form of it; nodebpy is the upstream answer.
    • GN evaluation returns no diff or summary: to know what the tree produced, the checker must evaluate the depsgraph and count (L-003).
    • Farm rendering needs a clean headless boot with no default cube and no UI-only settings (L-011).
    • Python edits push no undo step by default, so art-direction iterations are not transactional (L-008). Workaround: save versioned .blend files per accepted look.
  • Not feasible today:
    • Houdini-quality neon tube assets with physically plausible glass and gas glow at hero distance, without a human modeller’s eye.
    • EmberGen-speed smoke iteration; Mantaflow works but is slow at this scale.
    • 40 minutes at 11,200 × 2,160 on one workstation. Estimate: at 30 fps, 40 min ≈ 72,000 frames; at an assumed 2 GPU-minutes per frame that is ≈ 2,400 GPU-hours (interpretation; the source gives no fps or per-frame times).

5. Verification plan (RFC 0001 R6, six layers)

Section titled “5. Verification plan (RFC 0001 R6, six layers)”
Layer Check Threshold
1 Validity Every output EXR opens; resolution exact; required passes present; no NaN/Inf pixels 11,200 × 2,160 (or the declared test scale); passes {Combined, Emission, Denoising Albedo, Denoising Normal} present; 0 NaN/Inf
1 Validity GN evaluated geometry is valid 0 non-manifold edges on tube meshes; 0 degenerate faces
2 Spec Segment length and frame count Exactly the declared frames (e.g. 250 at 25 fps for 10 s); no gaps in the sequence
2 Spec Seamless loop Mean absolute difference between frame 0 and frame N+1 (rendered) < 0.5% of mean luminance
2 Spec No clipping on the LED wall < 0.1% of pixels with any channel > 1.0 after the delivery transform
2 Spec Temporal flicker Frame-to-frame mean luminance change < 2% except at declared cuts
3 Reference Determinism: same .blend, seed and frame rendered twice PSNR ≥ 45 dB; GN evaluated vertex positions identical (hash match)
3 Reference Tile stitch equals full render Max abs difference across the stitch seam < 1/255
4 Downstream Compositor reads the EXR; Emission + other light passes recombine to Combined Recombined vs Combined mean abs error < 1%
4 Downstream Delivery encode plays at wall resolution Encoded file decodes with exact frame count; bitrate within the venue’s spec (owner provides)
5 Appearance (warning) Vision model scores a 6-frame contact sheet against the brief (“Broadway neon, warm, readable from 20 m”) Warn if score < 7/10 or if text legibility is flagged
6 Taste (owner) Owner watches the loop on a scaled-down wall preview Owner sign-off; notes become parameter changes

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • The design language. What “Nashville” looks like as neon and distillery copper is a cultural and brand decision; Gentilhomme started in Illustrator for a reason.
  • Pacing at architectural scale. How slow is calm for a lobby where people stand for 30 seconds? That is judged in the room, not by a metric.
  • Hero assets. Neon lettering with believable glass, gas glow and bracket hardware needs a modeller’s eye; the agent can produce a clean first pass.
  • Sign-off with the venue. Brightness limits, content approvals and loop points are contractual.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, ~1 day): the owner sees, in their open Blender window, an empty scene become a wide orthographic frame filled with a grid of glowing tube segments. Pressing play, waves of brightness and displacement roll across the grid and loop after 10 s. Changing one group input (say speed or palette) on the GN modifier changes the whole wall. Three quarter-resolution EXR frames land in ~/newblender-data/demos/, with a check report like the chair’s.
  • Full reproduction: roughly 60–120 agent-hours to build 6 capsule systems and their checks; ≈ 2,000–3,000 GPU-hours of render (estimate above); 40–80 human-hours of direction (style frames, pacing reviews, venue sign-off), plus whatever hero assets are hand-modelled.

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • Frame rate, codec and LED processor constraints are not in the source; the checks above need the venue spec.
  • A vision model judging “neon mood” is unreliable (R6 layer 5 is a warning only).
  • Very wide frames stress GPU memory; border-split rendering adds a seam risk (hence the stitch check).
  • Our version of GN shading parity with Houdini neon is unproven.

9. Sources

Section titled “9. Sources”