DEGIS: Disentangled Palette Control for Brand-Coherent Image Generation

Digital advertising workflows increasingly use generative image models to produce visual content across campaigns, formats, and markets. However, reference-image conditioning entangles colour with subject identity, texture, and composition, making it difficult to control brand palettes independently. We introduce Disentangled Embeddings Guided Image Synthesis (DEGIS), a lightweight conditioning framework that separates colour palette control from spatial layout. A small multilayer perceptron maps frozen CLIP image embeddings to palette-specific embeddings, which are injected through a modified IP-Adapter, while Canny edges and ControlNet provide independent layout guidance. The diffusion backbone remains frozen, with a 3.5M-parameter Colour Head added for palette conditioning. On 240 paired advertising-layout evaluations held out from training, DEGIS improves palette matching over a whole-image IP-Adapter baseline, while preserving more layout and prompt information and reducing reference-identity leakage by approximately 45%. Expressed as colour gain per unit of reference identity copied, DEGIS achieves a colour-transfer efficiency of 0.72 [0.61, 0.86], compared with 0.33 [0.26, 0.40] for whole-image conditioning. These gains are achieved with approximately 4% additional latency and no additional memory overhead. The results demonstrate that disentangling palette information from whole-image reference conditioning provides more targeted palette control while preserving the flexibility of a frozen diffusion model.