Cluster Routed Identity Adaptation for Controllable Multimodal Content Generation Under Sparse Brand Supervision and Feedback Constrained Inference

• Vo Hung Cuong

  Faculty of Computer Science, Vietnam-Korea University of Information and Communication Technology, The University of Danang, Urban Area, Hoa Quy Ward, Ngu Hanh Son District, Da Nang 550000, Vietnam

  Author

• Pham Quang Huy

  Department of Software Engineering, University of Phan Thiet, 225 Nguyen Thong Street, Phu Hai Ward, Phan Thiet City 77000, Binh Thuan Province, Vietnam

  Author

• Dang Duc Long

  Department of Information Systems, Ha Tinh University, 447 26 Thang 3 Street, Dai Nai Ward, Ha Tinh City 450000, Ha Tinh Province, Vietnam

  Author

Generative content systems increasingly need to produce images that satisfy both semantic prompts and persistent identity constraints such as brand character, approved palette, recurring layout, and asset-specific visual language. General-purpose text-to-image models often preserve broad prompt semantics while drifting from these finer identity constraints, especially when the target identity is represented by only a few approved examples. This paper studies a cluster routed identity adaptation method for sparse content-identity generation. The method decomposes an identity repository into concept-level clusters, trains compact low-rank adapters for each cluster, routes a user prompt to one or more adapters through a joint text-image embedding space, and supplements prompts with brief-derived and caption-derived identity tokens when cluster support is limited. A second inference pass transfers masked background structure from approved reference assets while retaining generated foreground content. We evaluate the method on a simulated benchmark of 72 content identities, 11,840 approved assets, and 3,600 held-out generation prompts under two-shot, four-shot, eight-shot, and sixteen-shot regimes. Relative to a single identity adapter, the proposed method improves mean identity conformance from 0.673 to 0.761 in the four-shot setting, reduces palette error from 9.8 to 6.1 CIEDE2000 units, and lowers background FID from 38.4 to 29.7. Human preference judgments favor the routed method in 61.4\% of pairwise comparisons against retrieval-augmented prompting and 57.2\% against DreamBooth-style personalization. The results indicate that routing and low-data prompt supplementation can improve identity stability without fully retraining the base generator.

• PDF