Everything between the click and the buyer.

The 5-phase production system

Phases 00 and 03 halt for your approval. Phases 01, 02, and 04 run end-to-end. Nothing ships to your store without you signing off.

1. 00

   Calibration Halts for human review

   Eight anchor images, one per scene template. You approve the set. That set becomes the locked visual contract for every render after.

2. 01

   Inventory

   I read your Shopify catalog via API. Build a structured manifest: handle, title, type, dimensions, materials, scene-template match.

3. 02

   Generation

   Your brand constitution + the matched scene template + per-product subject block feeds into the image model. PNGs land in staging.

4. 03

   Human review Halts for approve / reject

   A review batch HTML pairs each staged image with its source product photo. You approve or reject per image, line by line.

5. 04

   Upload

   Approved scenes push to Shopify via GraphQL productCreateMedia. Appended to the gallery. Never replaces your primary catalog photo.

Brand constitution

A JSON spec I write with you in week one. It defines your brand visually: light, palette, architecture, props rules, composition, atmosphere, forbidden territory. Every batch after week one reuses it.

"brand": "Your brand",
"light": {
  "source": "single window natural",
  "kelvin": 3200–4200,
  "time": "late afternoon, golden hour",
  "forbidden": ["overhead flat", "strobe"]
},
"palette": {
  "cream": "#f5f0e6",
  "caramel": "#9c7a4e",
  "warm_brown": "#2a1e15",
  "forbidden": ["pure white", "jewel tones"]
},
"architecture": {
  "walls": "aged plaster, matte",
  "floor": "wide-plank aged oak",
  "anchor": "baseboard or doorway always visible"
}

The 99% fidelity engine

Six safeguards. Every batch runs through all six.

1. 01

   Multi-angle reference photos

   Your catalog photos passed as inline image data on every generation call, not just text descriptions.

2. 02

   Distinctive features array

   Per-product list of must-preserve details (hardware, joinery, finish, proportions, edge profile) injected into every prompt.

3. 03

   Subject proportions (5% tolerance)

   Width-to-height ratios computed from your dimensions string. Generation rejected if rendered ratio drifts more than 5%.

4. 04

   Scene scale anchors

   Every scene contains a known-size element: baseboard at 6 inches, door at 80 inches, window sash at 36 inches. Forces correct scale on the subject.

5. 05

   Vision QA pass

   After generation, a separate model scores the image on FEATURE, PROPORTION, and COLOR each 1–10. Below threshold, it auto-retries with corrective notes.

6. 06

   SHA256 spec hash

   Every approved image carries a content-addressable hash of the exact prompt, brand constitution version, and references that produced it. Audit trail.

Source → Rendered

Sixteen products from two real catalogs, one furniture brand on Shopify and one vanity brand on WooCommerce. Each row pairs the catalog product photo I sent the model as the fidelity anchor with the lifestyle scene the model returned. Same piece, every time. The hardware and joinery stay faithful to the source.

• Source

  Rendered

  Dawson entertainment console

• Source

  Rendered

  Cheltenham chair, blue velvet

• Source

  Rendered

  Keswick club chair, antique leather

• Source

  Rendered

  Reclaimed pine, double-arched bookcase

• Source

  Rendered

  Large cabinet, glass panel doors

• Source

  Rendered

  Garcia armchair

• Source

  Rendered

  York open-shelf washstand

• Source

  Rendered

  Tarbes tall mirror

• Source

  Rendered

  Rhone swivel chair, Belgian linen

• Source

  Rendered

  Winfall farmhouse-sink vanity

• Source

  Rendered

  French casement media console

• Source

  Rendered

  Hampden hand-carved arched mirror

• Source

  Rendered

  Lincoln carved double vanity

• Source

  Rendered

  Lori four-drawer chest

• Source

  Rendered

  Single vanity with drawers

• Source

  Rendered

  Salford serpentine chest

Ship to Google Merchant without flinching

The model layer applies the compliance rules before an image reaches review.

• Google Merchant 2026 ready

  Every PNG carries IptcExt:DigitalSourceType = trainedAlgorithmicMedia, the disclosure Google has required since 2024. The Q3 2026 C2PA-for-all-creative requirement is already baked in.

• SynthID + C2PA preserved

  Watermarking and content provenance baked in by the model, preserved end to end. Not stripped on upload.

• Alt-text disclosure

  Generated alt text discloses the image as AI-rendered. Shopify and Meta both accept this; it protects you on consumer-protection compliance.

• Never replaces your real product photo

  Renders push to the gallery only. Your primary catalog photo stays the real product. Lifestyle scenes earn the secondary slots.

• Named-designer prompt sanitizer

  Living designer names (the ones at risk in Andersen v. Stability AI) are stripped from prompts before they reach the model. Style without legal exposure.

• Content safety gate

  Auto-rejects faces, hands, text, logos, seals, alcohol, weapons. Catalog-safe by construction.