Hyperrealistic recolouring of existing photography

Client

Van de Velde

Van de Velde

Year

2025

2025

Sector

Marketing

Marketing

Problem

Van de Velde needs thousands of product images to support its international retailer network, its own retail chains and the marketing of its premium brands Marie Jo, Primadonna and Sarda. Successful lines regularly get new variants, which is why the group examined whether AI can extend existing photography in a controlled and high-quality way.

Solution

A custom Visual AI pipeline with custom-trained models that reuses existing photo shoot material and makes hyperrealistic packshots and lifestyle images possible, specifically for variants of existing products.

Result

More than 750 generated images at master resolution, delivered in 10 weeks from R&D to production. A scalable capacity that Van de Velde manages itself, trained on its own brands and embedded in the existing quality process.

Why AI image generation often fails in production

A convincing demo is one thing. A consistent premium brand catalogue at production scale is something else entirely.

AI image generation that works for a one-off marketing image does not automatically hold up once consistent, on-brand output is needed at production scale. Closed source models give limited insight into model behaviour or version management. Generic tools run into policy limitations, are hard to steer towards specific tasks and rarely reach the quality bar a premium brand demands. And ChatGPT or Google Gemini, however strong, are not designed for the degree of brand control Van de Velde needs.

That was the challenge for Van de Velde. The group has been building luxury lingerie since 1919 and is present internationally through a network of some 4,000 specialty stores and its own retail chains. For Marie Jo, Primadonna and Sarda, every image has to carry the same care, quality and brand consistency as the product itself. Where existing successful products get a new colour, existing imagery from classic photo shoots offers targeted opportunities for efficiency, without calling the role of physical shoots into question.

A convincing demo is one thing. A consistent premium brand catalogue at production scale is something else entirely.

AI image generation that works for a one-off marketing image does not automatically hold up once consistent, on-brand output is needed at production scale. Closed source models give limited insight into model behaviour or version management. Generic tools run into policy limitations, are hard to steer towards specific tasks and rarely reach the quality bar a premium brand demands. And ChatGPT or Google Gemini, however strong, are not designed for the degree of brand control Van de Velde needs.

That was the challenge for Van de Velde. The group has been building luxury lingerie since 1919 and is present internationally through a network of some 4,000 specialty stores and its own retail chains. For Marie Jo, Primadonna and Sarda, every image has to carry the same care, quality and brand consistency as the product itself. Where existing successful products get a new colour, existing imagery from classic photo shoots offers targeted opportunities for efficiency, without calling the role of physical shoots into question.

For Van de Velde, Visual AI is not a replacement for physical photo shoots but a targeted addition to existing image production. By reusing carefully selected photography for colour variants of proven products, we can work more efficiently without giving up quality, brand consistency or control.
How we approached it

We started with a research phase. R&D explored several routes, with particular attention to ethical use of AI and quality control. That led to a proof of concept: adapting packshots and marketing images of existing products to new fabrics, colours and textures. The goal was clear. Validate whether recolours meet Van de Velde's quality standards, and establish how much efficiency can be gained within the existing processes.

From there the pipeline took shape. We work with open source AI models, out of a choice for control, transparency, adaptability and a privacy-first approach to data. Models are trained for specific recolouring tasks, with an eye for the level of detail premium product images demand. Multiple models, parameters and processing steps are combined into a multi-step pipeline with QA gates, tuned to the way Van de Velde's teams work today.

The pipeline does not replace photo shoots. It supports a targeted extension of them. Existing photography forms the basis for colour variants of proven products, generated by custom-trained models that respect the look, feel and quality bar of each brand.

How we approached it

We started with a research phase. R&D explored several routes, with particular attention to ethical use of AI and quality control. That led to a proof of concept: adapting packshots and marketing images of existing products to new fabrics, colours and textures. The goal was clear. Validate whether recolours meet Van de Velde's quality standards, and establish how much efficiency can be gained within the existing processes.

From there the pipeline took shape. We work with open source AI models, out of a choice for control, transparency, adaptability and a privacy-first approach to data. Models are trained for specific recolouring tasks, with an eye for the level of detail premium product images demand. Multiple models, parameters and processing steps are combined into a multi-step pipeline with QA gates, tuned to the way Van de Velde's teams work today.

The pipeline does not replace photo shoots. It supports a targeted extension of them. Existing photography forms the basis for colour variants of proven products, generated by custom-trained models that respect the look, feel and quality bar of each brand.

From R&D to production in 10 weeks

In 10 weeks the track moved from R&D to a production-ready pipeline. The first delivery covered recoloured packshots and photo shoot images for colour variants of existing products, at master resolution and reviewed on colour-calibrated screens.

Alongside the technology itself, we advised Van de Velde on where Visual AI adds the most value: which existing product lines and colour variants take priority, how generated recolours fit within existing workflows, and how quality and governance stay under control as volume grows.

From R&D to production in 10 weeks

In 10 weeks the track moved from R&D to a production-ready pipeline. The first delivery covered recoloured packshots and photo shoot images for colour variants of existing products, at master resolution and reviewed on colour-calibrated screens.

Alongside the technology itself, we advised Van de Velde on where Visual AI adds the most value: which existing product lines and colour variants take priority, how generated recolours fit within existing workflows, and how quality and governance stay under control as volume grows.

What this has shown

This is what AI image generation can turn a promising prototype into: a reliable production capacity for controlled recolours. A custom pipeline that respects the look, feel and quality standard to the level where Van de Velde's existing quality process can build on it to guarantee a consistent body of imagery.

Book a meeting

Other work

text
Tomorrowland

Amare Social Media Experience

text
Kipling

Interactive Digital Christmas Card

text
Pukkelpop

Pukkelpop Demo

text
Tomorrowland

Amare Social Media Experience

text
Kipling

Interactive Digital Christmas Card

text
Pukkelpop

Pukkelpop Demo