Worn Checker Marble Floor

Project Overview

A seamless 4K black-and-white checker floor: scuffed, scratched, dusty stone with chipped joints and tiles that sit slightly proud or sunken. Like the chevron inlay, the Substance 3D Designer graph was written by Claude as a Python script and art-directed by me from renders.

This one tells a different story from the chevron. The chevron was a geometry problem: work out an exact pattern from a photo. A checker has no pattern to solve, so everything is in the surface. The first version came back as a clean veined marble that I rejected as looking stylized, and we rebuilt the material from scratch around wear instead of veining. The case study documents that turn and the mark-up loop that followed.

Role & Responsibilities

I art-directed the material and made every call on what “real” looked like: how scratched, how dusty, how uneven, how much shine survives. Claude wrote the bakers and the Designer script and diagnosed defects.

This case study is a pipeline demo first and a material second.

Key Contributions

  • Chose the material and the references, and switched the layout from a 45° diamond checker to square tiles
  • Rejected the first approach outright — the detailing read as stylized and the tile edges were far too clean
  • Rendered each iteration and reviewed it by eye against the references
  • Marked up screenshots with red arrows and one-line notes — “this swirling pattern looks strange, get rid of it”
  • Directed the wear one layer at a time: deeper scratches, dust, veins back in the black tiles only, lifted and sunken squares
  • Owns the Designer session, the Python source, the baked maps and the final renders

The Loop

Designer was never driven by hand. Every node, link, and value comes from a Python script run inside Designer, and the wear itself is baked outside Designer by a second script so it can be previewed before anything is run. I supply references, render, mark up, and decide. Claude bakes, generates, and diagnoses.

REFERENCE ANALYSE MODEL BAKE GENERATE RENDER REVIEW MARK-UP KNOB CHANGE · RE-RUN RE-BAKE FIX ME CLAUDE CLAUDE CLAUDE CLAUDE ME ME

Who Did What

Me

  • Chose the material, the references, and the bar for realism
  • Made the call to scrap the veined-marble approach and rebuild around wear
  • Rendered every iteration and reviewed each one against the references
  • Marked up renders with red arrows and one-line notes
  • Decided which wear goes where: veins and calcite spatter on black tiles only, dust everywhere
  • Owned the Designer session, the package, and the final renders

Claude

  • Wrote the first marble checker, then rebuilt the material around wear when it was rejected
  • Baked 14 maps at 4096² — joints, scratches, cuts, pits, veins, spatter, dust, per-tile lift and warp, roughness, 16-bit height
  • Wrote the Designer script that colours and combines the baked maps, with named knobs at the top of the file
  • Read my Designer install’s library list from the build log and wired in its grunge generators
  • Traced defects to their cause — ghost joint lines, a roughness range with no polish in it, 8-bit banding in the height
  • Kept everything re-runnable: bake script, graph script, one command each

Reading the Reference

Two references, used for different things. A game-environment screenshot of a scuffed black-and-white checker floor set the colour, the dust, and the scratch language. A second artist’s Substance checker floor, shown alongside its height and roughness maps, set the target for relief: tiles at different heights, each one slightly warped.

No measuring was needed this time. What the references gave was a list of things the first version did not have: scratches in several widths, scuff patches, pits, chipped and dirty joints, a dust film that lifts the blacks, and tiles that do not sit level.

Writing the Graph

The material is two scripts. A NumPy/OpenCV baker synthesises all the wear as fourteen 4096² maps, and an 853-line sd API script rebuilds the Designer graph around them on every run. Tuning lives at the top of the graph script as named constants: TILE_LIFT, UNEVEN, DUST_BASE, ROUGH_POLISH, VEINS, SPLAT and about twenty more.

Baking the wear outside Designer was a deliberate choice. Scratches, cuts, and chipped edges are drawn as thousands of individual strokes with full control over length, width, and clustering — and I can see the result before spending a Designer run on it.

Graph sections, in build order: Baked Wear Maps · Library Grunge · Base Color · Roughness · Height Normal AO · Outputs.

Stats: 853 lines · 14 baked maps · 6 framed sections · ~28 named knobs.

Ten of the fourteen baked wear maps synthesised in Python before Designer runs
Ten of the fourteen baked maps. Top: joints, pale marks, dark marks, veins, calcite spatter. Bottom: tone, roughness, height, per-tile warp, per-tile lift.

Render, Mark Up, Diagnose

I drove every change from a render and a sentence. After the heights were deepened, faint swirling rings appeared across the white tiles. I drew one arrow and wrote “this swirling pattern looks strange, get rid of it.” Claude didn’t blur them out. It identified them as contour lines: the smooth height data was being squeezed to 8 bits where the graph loaded it, so gentle slopes became tiny terraces, and the deeper relief had made them visible. The fix was to keep the height at 16 bits from file to normal map. Nothing in the look changed except the rings.

Four-panel diagnostic — markup, the per-tile warp at 8 bits with slopes exaggerated showing terraces, the same data at 16 bits, and the final normal map
My mark-up, the tile warp at 8 bits (slopes exaggerated, the rings are terraces), the same data at 16 bits, and the final normal map.

The rebuild followed the same loop. My note was that the detailing “almost looks more stylized and not as real world” and the tile edges were “way too clean.” Claude’s answer was that the look came from the marble recipe itself, and started again from wear.

Side by side — the first marble checker that was rejected and the rebuilt floor
The first marble checker I rejected, and the rebuilt floor.