Automating product imagery, saving millions in production costs.

The cost wasn’t taking the photograph. It was everything required to make it possible.

The studio had more than a thousand products to market, many selling for $30K–$40K. Buyers needed to see them inside real villas, hotels, courtyards and spas, not against a white background.

Each product needed around seven views: scale in a real room, architectural placement, material in daylight and at night, alternate angles and close detail. A 1,000+ product catalogue meant more than 7,000 finished images, and producing them conventionally meant fabrication, transport, installation, a location, styling, a crew and retouching every time.

One product reference in. A client-ready photography set out.

The studio supplies a sketch, render or source image. The system works out the product and its category, retrieves catalogue and client-preference context, designs an environment, decides realistic placement, generates one proof scene, evaluates it, locks the approved room, generates the rest of the shot family and packages everything for ecommerce.

• An interior design agent that plans the room before anything renders
• An image model that executes the brief
• A multimodal evaluator scoring fidelity, placement, room quality and taste
• Retrieval over approved and rejected references
• A production harness owning guardrails, retries, thresholds, routing and publishing

The first ten products proved the idea. Scaling it exposed what the model couldn’t decide alone.

Client review surfaced four failure classes. The model understood “luxury” generically but not this client’s taste. A sculpture could be inside a room without belonging there. Geometry, scale, materials and bases drifted. And seven views of “the same room” became seven different rooms.

At a hundred products, a rejected concept could invalidate a whole seven-image family. With early approval around 50%, hundreds of finished images could have no commercial value. Judgement had to move earlier than generation cost.

We stopped treating better prompts as the solution and built control systems around the models.

Approved and rejected references became structured client-preference memory covering lighting, greenery, architecture, furniture density, product dominance, placement and material palette. Product fidelity became a hard constraint, and one approved scene became the source of truth for every later view.

A concept eval runs before rendering, then a single proof image is scored before the remaining six views are produced. High-confidence concepts continue automatically; low-confidence ones go to a person. In a typical 500-product batch, roughly 450 continued automatically and around 50 needed review.