Our approach: distillation and dialogue
We worked with the studio to curate a high-quality dataset of Ross’ personal sketches, using it to fine-tune our text-to-image model, Imagen. By training the model on the studio’s selected work, we were able to incorporate the core components of Ross’ design language — the specific curves, structural logic and organic patterns, which allowed us to generate new concepts that were rooted in Ross’ unique style.
From the outset, we approached the project as a human-led inquiry. The studio determined the need to prioritize language alongside the visual dataset, working to decode and articulate Ross’ design lexicon to effectively guide the model’s output. We focused on building a specific vocabulary that described the studio’s work, knowing that the right prompts were key to getting meaningful results.
Throughout the process, Lovegrove Studio observed how the model responded to specific terms and used these insights to align the outputs toward the intended design outcomes. This dialogue between designer and AI was a crucial part of the process. We paid close attention to how the model interpreted certain words, using that feedback to refine prompts and steer the output closer to the studio’s vision. We challenged the model to generate a chair without ever using the word “chair,” instead using creative synonyms to produce more diverse outputs and a richer exploration of form and function.
The result: a physical prototype
We developed many concepts with this specialized model and the Lovegrove Studio team, then used Gemini to push the creative exploration further to ideate on materials and visualize the chair from different forms and viewpoints.
