acceptodds
Withdrawn from ICLR 2027

Personalized Image Generation with Reasoning and Reflection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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–render–reflect–re-render loop optimized with differential data reward. Across both tasks, PEARL substantially outperforms strong baselines, achieving an average improvement of 15% across personalization metrics.

Primary area:
generative models
Keywords:
Personalization, Language Model Applications

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BubpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies personalized image generation from users’ multimodal histories. It introduces PMH-IG, a benchmark covering two tasks: personalized scene generation using Amazon review histories and personalized creative generation using Instagram posting histories. The paper proposes PEARL, a framework that combines a multimodal planner and reflector with a frozen image generator. The planner infers user preferences and produces an initial generation prompt, while the reflector examines the generated image against the user’s history and revises the prompt.

Critical strengths
  1. The setup for the downstream applications is quite interesting.
  2. The paper is clearly written and easy to follow.
Critical weaknesses
  1. The paper does not report comparisons with the same Qwen2.5-VL directly generating prompts, training-free reflection, or best-of-N with the same number of image generations. It only compares against some other models.

  2. The results are only validated with SDXL, without testing on advanced image generation models. In addition, directly using the same model with some prompt engineering methods to expand prompts can also improve image generation quality.

  3. The user study has only one annotator, making it impossible to assess agreement across different people. Using one person’s preferences to judge whether images match other users’ histories is too subjective. Multiple annotators should be recruited, tests should be conducted using their own Instagram and Amazon information, and they should make the judgments themselves.

Questions for authors

See WeakNess

Reviewer_MSE7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduce PMH-IG, a benchmark comprising two tasks: (1) Personalized Scene Generation, which synthesizes product scenes reflecting user preferences inferred from Amazon review histories, and (2) Personalized Creative Generation, which generates images consistent with users' visual identities inferred from Instagram posting histories. The authors propose PEARL, a two-stage reasoning-and-reflection framework. A multimodal planner first extracts user preferences and produces a textual scene description, which is rendered by a frozen image generator. A reflector subsequently examines the generated image against the user's history and revises the scene description to improve personalization.

Critical strengths
  1. The paper addresses a meaningful limitation of traditional personalized image generation. I find this broader formulation of personalization interesting and potentially useful to the community.
  2. The proposed planner–reflector architecture provides an interpretable way to translate user histories into image-generation instructions.
Critical weaknesses
  1. Section 2.5 describes the MLLM-as-Judge protocol as a reference-free evaluation that measures alignment between generated images and users' historical preferences. However, Appendix F explicitly instructs the judge to compare the generated image against the user's held-out ground-truth post. These protocols measure fundamentally different properties: similarity to a specific reference image versus alignment with a user's historical preferences.
  2. The ablation studies do not isolate the contribution of reflection. The paper's central claim is that interleaving reasoning and reflection improves personalized image generation. However, the current ablation primarily compares the full PEARL framework against an SFT-only variant, which still includes the reflection stage. Consequently, the experiments evaluate the contribution of DPO rather than the reflection mechanism itself. Moreover, some metrics actually deteriorate after introducing DPO, such as Hit@5 and the overall MLLM score for Personalized Scene Generation. Moreover, the novelty of the reasoning-and-reflection mechanism is limited, similar iterative reasoning and visual feedback mechanisms have been explored in prior image-generation research, including IRG, which the paper itself discusses.
Questions for authors

refer to weakness

Reviewer_CBnPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors introduce a benchmark for personalized image generation from user histories, covering Personalized Scene Generation and Personalized Creative Generation. The proposed PEARL uses a planner, a frozen image generator, and a reflector to produce images aligned with a user's history. Experiments use Amazon review histories and Instagram post histories.

Critical strengths
  • The benchmark investigates personalization from accumulated user histories rather than a few reference images.
  • The complementary tasks and user-disjoint splits make this a useful setting for evaluating generalization to unseen users.
  • PEARL reports gains across several measures. In Table 2, inter-category user retrieval improves from 0.748 for Pigeon to 0.876 for PEARL, providing encouraging evidence for the complete pipeline.
Critical weaknesses
  • How were the Amazon and Instagram histories obtained, under what terms of use, how are user identities handled?

  • The scene-generation retrieval score does not isolate personalization from product preservation. Appendix D.2 ranks the generated scene against a user's history, although the target product is already supplied to the generator. A catalog-image baseline and a shuffled history control would help establish how much of the gain comes from the user's context.

  • The reflection ablation in Section 4.6 replaces preference optimization with supervised fine-tuning; it does not remove the reflector. It therefore cannot isolate reflection from the training change. The single-annotator human study offers limited independent support, and Appendix D.6 acknowledges different history windows for the creative-generation methods.

Questions for authors

Please address the points raised in the critical weaknesses section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.