ICG: Improving Cover Image Generation via MLLM-based Prompting and Personalized Preference Alignment

Zhipeng Bian; Jieming Zhu; Qijiong Liu; Wang Lin; Guohao Cai; Zhaocheng Du; Jiacheng Sun; Zhou Zhao

cs.CLApr 8, 2026

ICG: Improving Cover Image Generation via MLLM-based Prompting and Personalized Preference Alignment

Authors: Zhipeng Bian, Jieming Zhu, Qijiong Liu, Wang Lin, Guohao Cai, Zhaocheng Du, Jiacheng Sun, Zhou Zhao,

Organizations: Huazhong University of Science and Technology · Huawei Noah’s Ark Lab · Hong Kong Polytechnic University · Zhejiang University

Abstract

Recent advances in multimodal large language models (MLLMs) and diffusion models (DMs) have opened new possibilities for AI-generated content. Yet, personalized cover image generation remains underexplored, despite its critical role in boosting user engagement on digital platforms. We propose ICG, a novel framework that integrates MLLM-based prompting with personalized preference alignment to generate high-quality, contextually relevant covers. ICG extracts semantic features from item titles and reference images via meta tokens, refines them with user embeddings, and injects the resulting personalized context into the diffusion model. To address the lack of labeled supervision, we adopt a multi-reward learning strategy that combines public aesthetic and relevance rewards with a personalized preference model trained from user behavior. Unlike prior pipelines relying on handcrafted prompts and disjointed modules, ICG employs an adapter to bridge MLLMs and diffusion models for end-to-end training. Experiments demonstrate that ICG significantly improves image quality, semantic fidelity, and personalization, leading to stronger user appeal and offline recommendation accuracy in downstream tasks. As a plug-and-play adapter bridging MLLMs and diffusion models, ICG is compatible with common checkpoints and requires no ground-truth labels during optimization.

Figures & tables

Figure 1: The overall pipeline for cover image generation.

Figure 2: Overview of the proposed method. The model takes a reference image, title, and meta token to generate context embeddings via a Multimodal LLM. Combined with user embeddings, personalized features are injected into a diffusion model through a dual-path adapter. Reward models evaluate the output and guide training via feedback.

Figure 3: Model designed to train meta token.

Dataset	PixelRec				MovieLens
Metric	LPIPS( $\downarrow$ )	SSIM( $\uparrow$ )	FID( $\downarrow$ )	Aesthetics( $\uparrow$ )	LPIPS( $\downarrow$ )	SSIM( $\uparrow$ )	FID( $\downarrow$ )	Aesthetics( $\uparrow$ )
Title+Image Rule-based	0.6446	0.1484	47.74	4.17	0.6512	0.1634	46.24	4.09
Text Inversion gal2022image	0.6282	0.1632	42.23	4.12	0.6345	0.2474	43.27	4.12
PMG PMG	0.5411	0.1624	35.18	4.21	0.4140	0.2515	33.93	4.11
ICG	0.5126	0.1724	33.06	4.87	0.4018	0.2695	31.23	4.77

Table 1: Quantitative comparisons. The best results are in bold and the second-best results are underlined .

Figure 4: Qualitative comparison. Content titles, reference images and generated covers with different approaches

	PixelRec	MovieLens
ICG	2.419	2.527
Title+Image Rule-based	1.978	2.041
Text Inversion	1.952	1.923
PMG	2.152	1.994

Table 2: The average score of generated covers in human evaluation

$N$	Finetuning	PixelRec	MovieLens
1	✗	0.4367	0.5491
2	✗	0.4359	0.5482
4	✗	0.4398	0.5526
8	✗	0.4495	0.5689
1	✓	0.4194	0.5293
2	✓	0.4168	0.5315

Table 3: Quantitative ablation study of multimodal generative learning stage and meta tokens using the LPIPS metric two datasets.

N

denotes the number of multimodal tokens. The best results are in bold and the second-best results are underlined .

Figure 5: The effectiveness ablation of varying user conditions.

Dataset	PixelRec		MovieLens
Metric	LPIPS( $\downarrow$ )	FID( $\downarrow$ )	LPIPS( $\downarrow$ )	FID( $\downarrow$ )
ICG	0.5126	33.06	0.4018	31.23
w/o Meta token	0.5912	39.24	0.5854	37.02
w/o User feature	0.5203	32.67	0.4284	31.43
w/o Both	0.5893	38.45	0.5194	36.54

Table 4: Overall quantitative ablation study of the ICG framework. The best results are in bold and the second-best results are underlined .

Figure 6: The effectiveness ablation of the proposed CLIP, PickScore and HPSv2 rewards.

Metric	LPIPS( $\downarrow$ )	FID( $\downarrow$ )	Aesthetics( $\uparrow$ )
ICG	0.5126	33.06	4.87
w/o CLIP	0.5504	35.76	4.71
w/o HPSv2+PickScore	0.5413	34.87	4.45
w/o Personalized Reward	0.5653	35.81	4.54

Table 5: Quantitative ablation study of the reward models. The best results are in bold and the second-best results are underlined .

Figure 7: Generated Example Covers. Despite being trained on the base Stable Diffusion v1.5, our model can be seamlessly applied to a range of community checkpoints.

	PixelRec	MovieLens
Personalized Reward Model	85.2	86.2
Only image	53.8	54.1
Image and title	61.3	67.1
Image and user profile	74.6	78.3
w/o transformers	70.5	72.5

Table 6: Personalized preference prediction accuracy on test sets of PixelRec and MovieLens under different setting

	Item	User	Recall@10	NDCG@10
w/o image	✗	✗	16.17%	0.0749
Item	✓	✗	17.94%	0.0853
Averaged-user	✓	Average	18.99%	0.0991
Generated-user	✓	Generated	20.21%	0.1016

Table 7: Comparison of MMGCN’s recommendation performance using different item and user image features. Best results are highlighted in bold, and second-best results are underlined.

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 461 papers

Personalize at Test Time: Learning User Preferences for Image Generation
Oct 6, 2026Jiamu Bai, Jiaming Hu, Yanhong Wu +1User Preference Modeling Diffusion Model Alignment
Penn State University · Meta · Boston University +1
Personalized Image Generation with Reasoning and Reflection
Sep 30, 2026Bo Ni, Ngoc N. Tran, Qinwen Ge +10Image Generation Agentic Image Generation
Vanderbilt University · Adobe Systems · University of Maryland, College Park +1
Polaris: Scaling Up Instruction-Guided Image Generation Towards Millions of Personalized Style Needs
Jun 1, 2026Zhi-Kai Chen, Jun-Peng Jiang, Jun-Jie Tao +2Image Generation Controllable Image Generation
School of Artificial Intelligence, Nanjing University, China · National Key Laboratory for Novel Software Technology, Nanjing University, China · Nanjing University, China

Oct 6, 2026J/K move · Enter open · S save

Personalize at Test Time: Learning User Preferences for Image Generation

Jiamu Bai, Jiaming Hu, Yanhong Wu, Zellux Wang

Penn State University · Meta · Boston University+1

User Preference Modeling Diffusion Model Alignment Image Generation Reward Modeling Diffusion Model Guidance

Open paper AskPDF arXiv

Diffusion models can generate high-quality images, yet aligning their outputs with individual user preferences remains challenging. A key bottleneck is accurately modeling diverse user preferences from limited feedback. Existing approaches often rely on labor-intensive manual preference annotations or vision-language models (VLM) to extract preference information from user interaction histories, introducing substantial annotation or computational costs that limit scalability. We propose an approach that learns personalized reward models directly from users' historical image preference pairs. First, we use an autoencoder to compress hundreds of visual attributes into 50 attribute-anchored preference dimensions and train an evaluator to score images along these dimensions. We then represent each user's preferences as a linear combination of the shared dimension scores, estimating the user-specific weights by maximizing the likelihood of their observed pairwise preferences under the Bradley-Terry model. This formulation reduces per-user adaptation to optimizing a low-dimensional weight vector, simplifying optimization and enabling data-efficient personalization from sparse feedback. The learned personalized rewards guide image generation at inference time while keeping the diffusion model frozen. Experiments on real-user preference data show that our approach achieves approximately 77% held-out pairwise preference prediction accuracy and improves the alignment of generated images with individual user preferences.

Abstract

Image Generation

Explore similar work

Personalize at Test Time: Learning User Preferences for Image Generation

Personalized Image Generation with Reasoning and Reflection

Polaris: Scaling Up Instruction-Guided Image Generation Towards Millions of Personalized Style Needs