Personalized Image Generation with Reasoning and Reflection

Bo Ni; Ngoc N. Tran; Qinwen Ge; Franck Dernoncourt; Seunghyun Yoon; Samyadeep Basu; Sungchul Kim; Puneet Mathur

cs.CVSep 30, 2026

Personalized Image Generation with Reasoning and Reflection

Authors: Bo Ni, Ngoc N. Tran, Qinwen Ge, Franck Dernoncourt, Seunghyun Yoon, Samyadeep Basu, Sungchul Kim, Puneet Mathur, +5 more

Organizations: Vanderbilt University · Adobe Systems · University of Maryland, College Park · University of Georgia

Abstract

Personalized image generation has remained narrowly focused on conditional synthesis from curated visual exemplars, rather than capturing who a user is. In practice, however, a user's personal context is much richer, comprising reviews, posts, images, captions, and metadata accumulated over time. A truly personalized generator should leverage this history to produce images aligned with the user's lifestyle and aesthetic preferences. To this end, we introduce the first unified benchmark for personalized image generation from user histories. The benchmark comprises two complementary tasks and a multi-axis evaluation protocol that assesses target fidelity, visual quality, user distinguishability, semantic alignment with the user's history, and task-specific utility. Grounded in real-world e-commerce and social media settings, the benchmark includes: (1) Personalized Scene Generation, which places a given object in a scene that reflects a user's preferences and lifestyle, motivated by personalized product presentation; and (2) Personalized Creative Generation, which generates a novel image on a specified topic that is faithful to a user's aesthetic and visual identity, motivated by social media content creation. We further propose PEARL, which couples a multimodal reasoner with a frozen image generator in an interleaved reason-reflect loop optimized with differential data reward. Across both tasks, PEARL outperforms strong baselines, achieving an average improvement of 15% across personalization metrics.

Figures & tables

Figure 1: Personalized Image Generation based on User History.

Figure 2: Training procedure of the Pearl framework. We warm start with the Silver Personalization Trajectory, and optimize both policies with the Render-in-the-Loop optimization.

	Retrieval		Image Quality	MLLM Judge ( $1$ – $5$ )
Method	H@5 $\uparrow$	MRR $\uparrow$	Aes $\uparrow$	Cont. $\uparrow$	Overall $\uparrow$	Sty. $\uparrow$
PMG ( Shen et al., 2024 )	0.2143	0.1535	5.11	3.58	3.48	2.48
Pigeon ( Xu et al., 2025 )	0.0985	0.0834	4.60	4.32	1.79	1.98
LaVIT ( Jin et al., 2024 )	0.2162	0.1441	4.38	3.94	2.92	2.37
LLaVA ( Liu et al., 2023 )	0.1873	0.1367	5.17	4.32	2.89	2.30
Pearl (ours)	0.2297	0.1639	5.33	4.36	3.92	2.88

Table 1: Main results on Personalized Scene Generation for e-commerce setting on Amazon. (Bold marks the best per column; underline marks second-best.)

	Target-based Image Metrics					Contrastive R@1		MLLM Judge ( $1$ – $5$ )
Method	CIS $\uparrow$	DIS $\uparrow$	LPIPS $\downarrow$	MS-SSIM $\uparrow$	Aes. $\uparrow$	Inter-cat. $\uparrow$	Intra-cat. $\uparrow$	Sty. $\uparrow$	Cont. $\uparrow$	Overall $\uparrow$
PMG ( Shen et al., 2024 )	0.601	0.206	0.729	0.047	5.84	0.597	0.466	3.167	3.075	3.157
Pigeon ( Xu et al., 2025 )	0.561	0.256	0.741	0.053	4.28	0.748	0.659	3.341	3.258	3.334
LaVIT ( Jin et al., 2024 )	0.616	0.238	0.744	0.068	5.37	0.640	0.395	3.344	3.143	3.331
LLaVA ( Liu et al., 2023 )	0.575	0.171	0.735	0.053	5.48	0.432	0.285	3.256	3.179	3.227
Pearl (ours)	0.645	0.289	0.735	0.055	5.45	0.876	0.701	3.554	3.277	3.541

Table 2: Main results on Personalized Creative Generation for social media setting on Instagram. (Bold marks the best per column; underline marks second-best.)

Panel of Figure 3 — Figure 3: Qualitative Results.

Figure 4: Ablation: Pearl vs. Pearl -Reflection. The visualization normalizes and aggregates the metrics across both tasks.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

	Personalized Scene Generation (E-commerce: Amazon)		Personalized Creative Generation (Social Media: Instagram)
Split	#Users	#Reviews	#Users	#Posts	#Pairs
Train	3,972	39,720	3,464	40K	957
Val	510	5,100	742	8K	—
Test	518	5,180	207	2K	314
Total	5,000	50,000	4,413	50,656	1,271

Appendix

Table 3: Dataset statistics.

Task	Criterion	Pearl	Baseline	Equal	Other
Scene	History fit	19 (76%)	4 (16%)	0 (0%)	2 (8%)
Scene	Product preservation	8 (32%)	3 (12%)	9 (36%)	5 (20%)
Creative	History fit	16 (64%)	9 (36%)	0 (0%)	0 (0%)
Creative	Topic adherence	18 (72%)	5 (20%)	2 (8%)	0 (0%)

Appendix

Table 4: Human evaluation of the original SDXL outputs. Entries are counts (percentages), with 25 pairs per task and question. The baseline is PMG for scenes and Pigeon for creative generation. “Other” denotes insufficient evidence for history fit and neither image satisfactory for product preservation or topic adherence.

	Retrieval		Image Quality	MLLM Judge ( $1$ – $5$ )
Method	H@5 $\uparrow$	MRR $\uparrow$	Aes $\uparrow$	Cont. $\uparrow$	Overall $\uparrow$	Sty. $\uparrow$
Pearl -Reflection	0.2336	0.1583	5.32	4.14	3.99	2.74
Pearl (ours)	0.2297	0.1639	5.33	4.36	3.92	2.88

Appendix

Table 5: Ablation Study for Personalized Scene Generation. Bold marks the best per column.

	Target-based Image Metrics				Contrastive R@1		MLLM Judge ( $1$ – $5$ )
Method	CIS $\uparrow$	DIS $\uparrow$	LPIPS $\downarrow$	MS-SSIM $\uparrow$	Inter-cat. $\uparrow$	Intra-cat. $\uparrow$	Sty. $\uparrow$	Cont. $\uparrow$	Overall $\uparrow$
PEARL-Reflection	0.635	0.285	0.732	0.057	0.838	0.701	3.529	3.331	3.516
Pearl (ours)	0.645	0.289	0.735	0.055	0.876	0.701	3.554	3.277	3.541

Appendix

Table 6: Ablation Study for Personalized Creative Generation. Bold marks the better of the two per column.

Panel of Figure 5 — Figure 5: Additional qualitative results on Personalized Creative Generation. For each of three users, the top row shows five historical posts alongside a held-out target post (rightmost), and the bottom row shows generations from LaVIT, LLaVA, PMG, Pigeon, and Pearl . The examples illustrate recurring choices in color, framing, and scene presentation across interiors, food, and quilting content.

Symbol	Description
Problem Setup
$u$	A user.
$\mathcal{H}_{u}$	User $u$ ’s multimodal history; a sequence of prior activities.
$h^{(u)}_{i}$	The $i$ -th history entry for user $u$ (e.g., a review with product image, or a post with caption).
$N_{u}$	Number of history entries for user $u$ .
$T$	A personalized generation task.

Appendix

Table 7: Notation used throughout the paper.

Explore similar work

PIPBench: A Profile-Inclusive Framework for Personalized Image Generation Evaluation
Jul 7, 2026Yuhang Wu, Shuxiang Zhang, Wee Hian Ching +2Aesthetics Synthetic Data
College of AI, Tsinghua University · Sun Yat-sen University · Shanghai Qi Zhi Institute
ZIPP:Zero-shot Image Personalization from Personas
Jun 7, 2026Harini SI, Somesh Singh, Yaman Kumar Singla +2Persona Consistency Diffusion Models
Adobe Media and Data Science Research (MDSR) · SUNY at Buffalo · IIIT-Delhi
Polaris: Scaling Up Instruction-Guided Image Generation Towards Millions of Personalized Style Needs
Jun 1, 2026Zhi-Kai Chen, Jun-Peng Jiang, Jun-Jie Tao +2Multi-Reference Image Generation Image Generation
School of Artificial Intelligence, Nanjing University, China · National Key Laboratory for Novel Software Technology, Nanjing University, China · Nanjing University, China

Jul 7, 2026, cs.CVJ/K move · Enter open · S save

PIPBench: A Profile-Inclusive Framework for Personalized Image Generation Evaluation

Yuhang Wu, Shuxiang Zhang, Wee Hian Ching, Chi Zhang+1

College of AI, Tsinghua University · Sun Yat-sen University · Shanghai Qi Zhi Institute

Aesthetics Synthetic Data Prompt Engineering

Open paper AskPDF arXiv

Recent text-to-image models such as DALLE-3 excel at following diverse prompts yet remain blind to individual aesthetic preferences. We study personalized image generation, where models must align outputs with a user's implicit visual preferences based on a few historically preferred images and a short prompt. To this end, we introduce PIPBench, the first profile-inclusive benchmark for evaluating personalized image generation. We further propose a novel data construction pipeline that leverages psychological and demographic profiling dimensions for both real-user data collection and scalable agent-based data generation. Using PIPBench, we conduct a thorough evaluation of representative line of methods. Our experiments reveal key limitations in existing methods, suggesting new challenges and opportunities for personalized text-to-image synthesis. Project page: https://wuyuhang05.github.io/PIPBench/

Abstract

Explore similar work

PIPBench: A Profile-Inclusive Framework for Personalized Image Generation Evaluation

ZIPP:Zero-shot Image Personalization from Personas

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