User preference prediction requires a comprehensive and accurate understanding of individual tastes. This includes both surface-level attributes, such as color and style, and deeper content-related aspects, such as themes and composition. However, existing methods typically rely on general human preferences or assume static user profiles, often neglecting individual variability and the dynamic, multifaceted nature of personal taste.

To address these limitations, we propose an approach built upon Multimodal Large Language Models, introducing contrastive preference loss and preference tokens to learn personalized user preferences from historical interactions. The contrastive preference loss is designed to effectively distinguish between user "likes" and "dislikes", while the learnable preference tokens capture shared interest representations among existing users, enabling the model to activate group-specific preferences and enhance consistency across similar users.

Extensive experiments demonstrate our model outperforms other methods in preference prediction accuracy, effectively identifying users with similar aesthetic inclinations and providing more precise guidance for generating images that align with individual tastes.

Figure 2: Overview of our MLLM-based preference learning framework.

(a) The visual encoder and text embedding module extract preference representations \( x_u^{+/-} \) by processing the preference history \( \mathcal{S} \) and a target item \( z_{\text{pos/neg}} \).

(b) The framework is trained using a base loss \( L_{\text{base}} \) to predict preference labels, and a contrastive preference loss \( L_{\text{CP}} \) that enhances separability between liked and disliked items. Additionally, learnable preference tokens \( P_v \) are introduced to model shared user interests.

Figure 3: Mining Similar Users via Attention Mechanism.

(a) Attention scores \(\mathcal{A}\) represent interactions between preference tokens and target image tokens for individual users. Each user has a unique reference history, and we concatenate the same target image to the input sequence across users. For each user, the horizontal axis represents tokens from the target image, while the vertical axis represents the preference tokens. Each user has five different random re-orderings of reference images.

(b) Examples of images liked (✓) or disliked ( × ) by each user.