WallMock — Catalog Director

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for WallMock — Catalog Director

Inspiration

High-volume wall-art sellers repeat the same production work for every listing: choose a room, decide whether the ratio fits, size the piece, place it relative to the furniture, choose a frame, export several channels, and repeat. Existing visualization tools are useful editors, but a seller with hundreds of designs still makes most of those decisions manually.

WallMock asks a different question: what if the seller uploads a collection once and software directs the whole visual-merchandising job?

What it does

WallMock analyzes each artwork, ranks certified rooms, solves a credible physical size and placement, applies the selected product construction, and composites the seller's exact pixels into the scene. It verifies fidelity and geometry before an output can be delivered. The working app generates four mockups, reports real scale and exact-art scores, and downloads individual images or a ZIP.

Catalog Director is the Build Week extension. It plans a consistent five-role listing pack—Hero, Lifestyle, True Scale, Product Detail, and Options Card—across an entire collection. It carries seller preferences and collection intent, preserves visual diversity, promotes ordered fallbacks safely, resumes jobs idempotently, and withholds anything that lacks the required spatial, rights, or release authority.

How we built it

The image engine is TypeScript on Node.js with Sharp/libvips. Room geometry, homographies, frame construction, shadows, product footprints, and fidelity checks are deterministic and test-pinned. Generative models can help author room sources, but the final artwork is never redrawn.

Codex with GPT-5.6 was the engineering orchestrator and adversarial reviewer for the Build Week work. It mapped a large pre-existing codebase, reconciled the product vision with contradictory legacy behavior, drafted executable milestone specs, implemented and reviewed fail-closed catalog authorities, diagnosed geometry and numerical bugs by running the real pipeline, generated law tests, and inspected visual evidence. The process intentionally used disagreement: one pass built a change and another tried to refute it before owner approval.

Challenges

The hardest problem was not pasting an image onto a wall. It was knowing when a result was untrustworthy. AI-authored rooms can contain ambiguous openings, painted frame residue, broken wrap sides, or objects whose depth order is wrong. Several attempted repairs passed simple metrics while still looking wrong at native zoom.

We changed the system to measure and withhold. Frame openings are fitted once, product footprints are rechecked after every translation, crop survival is proven before rendering, True Scale requires measured or declared PPI, and owner-rejected source rooms are retired instead of receiving room-specific exceptions.

Accomplishments

• Four real mockups generated in 19.2 seconds in the recorded demo.
• Exact-art SSIM scores of 0.9512, 0.9651, 0.9669, and 0.9716.
• Fail-closed commercial authority and catalog certification.
• Role geometry for crop-safe Hero, truthful True Scale, and coded Product Detail.
• Resumable collection planning with diversity-constrained fallbacks and deterministic artifact IDs.
• A review trail that records rejected sources and preserves the live catalog until explicit owner approval.

What we learned

AI makes code and source imagery cheaper, but reliability still comes from accumulated decisions, measured constraints, customer corrections, and trust. The strongest use of GPT-5.6 was not a single prompt. It was a disciplined loop of repository analysis, implementation, execution, refutation, visual review, and evidence-backed acceptance.

What's next

WallMock will expose Catalog Director in the seller UI, add secure managed billing and durable jobs, and capture privacy-safe feedback such as accepted placements, room replacements, and size changes. A learned ranker will then choose among candidates that already pass physical and product laws. With seller consent, the system can also learn which visual roles improve engagement and conversion.

Longer term, sellers will be able to save brand-specific frame profiles, create private premium room packs, connect Shopify/Etsy catalogs, and use WallMock as a placement and visualization API. The compounding asset is not a prompt or a room count; it is the structured knowledge of which products, placements, and visual stories work for which art collections.

Built With

• codex
• node.js
• openai
• playwright
• sharp
• typescript
• vitest