Google DeepMind fine-tuned Imagen on one designer's sketches to make a 3D-printed chair

A single-studio fine-tune, not a general design tool

Google DeepMind's project with Lovegrove Studio, Creative Director Ila Colombo, and design office Modem is narrower than the headline suggests. The team fine-tuned Imagen, its text-to-image model, on a curated dataset of Ross Lovegrove's personal sketches. The output is not a design assistant for everyone — it's a model built to reproduce one designer's specific curves, structural logic, and organic patterns.

The chair itself was chosen deliberately. Google frames it as a classic constraint problem: a fixed function paired with an open form. That framing let the team measure whether the model was expressing Lovegrove's style rather than just producing plausible furniture. The success criterion was subjective — whether the studio felt the result was an authentic extension of their work — which is a different bar than accuracy or benchmark scores.

The real engineering was a shared vocabulary

The most transferable detail is that the studio decided to prioritize language alongside the visual dataset. They worked to decode and articulate Lovegrove's design lexicon — building a specific vocabulary to describe the studio's work — because the visual training alone wasn't enough to steer the model toward intended outcomes.

The team then treated prompting as an iterative loop: they watched how the model interpreted particular terms and used that feedback to refine prompts. The sharpest example is that they challenged the model to generate a chair without ever using the word 'chair,' substituting synonyms to force a wider exploration of form. That's a concrete prompt-engineering tactic, and it reveals how much the interface between designer and model was a linguistic one, not a slider or a UI.

Two models doing two jobs

The workflow used two distinct systems for two distinct stages. The fine-tuned Imagen model produced concepts in Lovegrove's style. Then Gemini was used downstream to ideate on materials and visualize the chair from different forms and viewpoints. The generation and the exploration were handled separately, which is worth noting for anyone assuming a single model can carry a full design pipeline.

The pipeline ended in atoms, not pixels. The final design was produced physically using metal 3D printing. Google describes this as transforming AI-generated pixels into a tangible piece — the demonstration is that the outputs held up through manufacturing, not just on screen.

What a per-artist fine-tune implies for creative tooling

The lesson from this specific project is that capturing a personal style required both a private, curated dataset and a human-authored vocabulary to drive it. Neither the base model nor a generic prompt would have reproduced Lovegrove's language; the studio had to teach the system its own terms and observe the responses to close the gap.

For me, the final result transcends the whole debate on design. It shows us that AI can bring something unique and extraordinary to the process. — Ross LovegroveMontana Labs

For applied teams building creative tools, the takeaway is that the durable asset here is the paired dataset-and-lexicon per collaborator, not a one-size model. The frontend of a system like this is a shared descriptive language, and building it is a slow, human-led curation task — the part that doesn't come for free with the model.