Context-aware intelligent colour design via multi-objective optimisation and vision–language models

Selecting a colour palette that suits its design context (application domain, audience, and accessibility profile) affects usability, engagement, and accessibility, yet remains a difficult computational problem. Vision–language models (VLMs) have improved performance on this task but share three limitations: deterministic outputs cannot represent the variance inherent in a design brief, single-scalar training fixes the harmony–accessibility trade-off at training time, and post-hoc context filtering leaves the generator itself context-blind. We argue that these limitations stem from the output representation rather than the encoder. CAIM (Context-Aware Intelligent palette Modelling) represents palettes as a context-conditional Gaussian mixture in CIELAB whose parameters are predicted from frozen vision–language embeddings by a learned head. It refines sampled candidates under five objectives designed to target distinct and complementary mathematical features with reference-direction-based many-objective optimisation and ranks them with a context-conditioned selection rule. On four public datasets and a 3200-scenario context-rich dataset, CAIM achieves up to 23% higher predicted ratings than the strongest baseline, a 22-percentage-point gain in APCA readability over the strongest baseline, and a Pareto hypervolume 34% larger than the convex hull of baseline solutions, at 1.6 s per scenario. Modelling VLM-conditioned generation as a context-conditional distribution is promising for design tasks with inherently uncertain outputs.