Pith. sign in

REVIEW 4 major objections 9 minor 2 cited by

More reference images hurt personalized image generation

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-08 05:35 UTC pith:6MAGXBTA

load-bearing objection First profile-inclusive benchmark for personalized image generation; real-user data is the strength, candidate-generation pipeline creates a structural circularity that the paper does not fully address. the 4 major comments →

arxiv 2607.06440 v1 pith:6MAGXBTA submitted 2026-07-07 cs.CV

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

Yuhang Wu , Shuxiang Zhang , Wee Hian Ching , Chi Zhang , Miao Liu This is my paper
classification cs.CV
keywords personalized image generationuser preference modelingpsychological profilingbenchmarkvision-language modelstext-to-image generationsynthetic agentspreference alignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces PIPBench, the first benchmark for personalized image generation that pairs real user profiles—psychological traits, demographics, lifestyle—with their preferred images. The benchmark comprises 1,369 test cases drawn from 76 real users and 175 synthetic agents. Using this benchmark, the authors evaluate four paradigms for conditioning image generation on user preference: test-time tuning, joint conditioning, condition fusion via VLM prompt enrichment, and separate conditioning. The central finding is that VLM-based prompt enrichment—where a vision-language model rewrites a short user prompt by reasoning over reference images—substantially outperforms direct joint conditioning approaches (GPT-5 Elo: 1765 vs. 1521 for single-reference joint conditioning on real-user data). More strikingly, providing two reference images to a joint conditioning model degrades performance compared to providing one, because current image generation models cannot effectively interpret multiple reference images simultaneously. The paper also demonstrates that synthetic agents constructed via psychologically grounded profile sampling produce training data that transfers to real-user preference alignment, and that profile-conditioned data generation yields more diverse and coherent preference sets than profile-free tag-based generation.

Core claim

The paper's central discovery is that the dominant strategy for personalized image generation is not direct multi-image conditioning but rather indirect textual mediation: a vision-language model reads multiple reference images, infers the user's aesthetic preferences, and rewrites the generation prompt accordingly. This VLM fusion approach achieves the highest preference alignment scores. In contrast, joint conditioning models that attempt to inject multiple reference images directly into the generation process fail to benefit from additional references—two references perform worse than one—revealing that current image generation architectures lack the capacity to jointly reason over multi-

What carries the argument

The paper builds on a profile-to-aesthetic mapping pipeline: psychological frameworks (Big Five Openness, Schwartz Values, Ecological Valence Theory, Circumplex Model of Affect) and personal context variables (academic background, lifestyle, hobbies, fashion) are encoded into structured profiles. An LLM then performs chain-of-thought reasoning over these profiles to generate profile-conditioned image prompts. For synthetic agents, profiles are stochastically sampled from a schema and validated through hard consistency rules (rejecting logical contradictions like high curiosity with high rigidity) and soft conflict scoring, with Jaccard diversity filtering ensuring heterogeneity. The persona-

Load-bearing premise

The benchmark's synthetic component (719 of 1,369 test cases) assumes that LLM-generated psychological profiles and LLM-reasoned mappings from personality traits to visual aesthetics accurately reflect real human visual preferences, without independent empirical validation that these specific trait-to-aesthetic mappings hold for image generation.

What would settle it

If a model trained on profile-free synthetic data matched or exceeded a model trained on profile-inclusive synthetic data on real-user preference alignment, the claim that psychological profiling improves preference data quality would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Image generation architectures need redesigned conditioning mechanisms that can jointly attend to and synthesize signals from multiple reference images, rather than treating them as independent style inputs.
  • The finding that textual mediation outperforms direct visual conditioning suggests that the bottleneck in personalized generation is preference interpretation, not generation fidelity.
  • Synthetic agents grounded in psychological profiles can serve as scalable proxies for underrepresented user populations in preference evaluation, though the validity of this proxy depends on the accuracy of profile-to-aesthetic mappings.
  • The persona-aware Elo evaluation protocol—where an LLM judge adopts a user's profile to compare generated images—achieves ~91% agreement with human annotators, offering a reproducible alternative to costly human preference studies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If VLM prompt enrichment already outperforms end-to-end visual conditioning, the performance gap may stem from the diffusion model's cross-attention mechanism being architecturally unsuited for multi-image preference extraction, rather than from a fundamental limitation of the conditioning paradigm. A model that explicitly performs cross-image reasoning before conditioning might close this gap.
  • The observation that real users are easier to align than synthetic agents could partly reflect that real-user preferences were themselves shaped by internet-distributed AI-generated imagery, creating a circularity: models trained on web data already encode the aesthetic distribution from which these users form their preferences.
  • The profile-to-aesthetic mappings (e.g., high Openness → abstract compositions) are operationalized entirely through LLM reasoning without independent empirical validation for image generation contexts. If these mappings are inaccurate or culturally specific, the synthetic agent benchmark may measure LLM aesthetic reasoning consistency rather than genuine human preference structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. PIPBench introduces a profile-inclusive benchmark for personalized image generation that pairs user profiles (psychological traits, demographics) with preferred images. The benchmark combines a real-user dataset (76 users, 650 test cases) with a synthetic agent dataset (175 agents, 719 test cases). The pipeline uses LLM-based chain-of-thought reasoning over psychological frameworks (TIPI, Schwartz, EVT, CMA) to generate profile-conditioned image prompts, from which candidate images are produced and preferences are elicited. The authors evaluate representative methods across four paradigms (test-time tuning, joint conditioning, condition fusion, separate conditioning) using both automatic metrics and a persona-aware Elo rating system with LLM-as-a-judge. The headline finding is that VLM-based condition fusion (GPT-5) achieves the best preference alignment (Elo 1765 on real-user data), while joint-conditioning methods degrade when given more reference images. The paper also proposes learnable approaches (VLM instruction tuning, preference compression) and validates the profile-inclusive framework against a profile-free baseline.

Significance. The paper addresses a genuine gap: existing personalized image generation benchmarks lack holistic user profiles and rely on isolated cues (single reference images, prompt histories). The combination of real-user data with psychological profiling and a scalable synthetic agent pipeline is a reasonable design. The persona-aware Elo rating achieves ~91% human agreement on a 200-case subset, and a 100-case user study (Table 6) confirms the ranking. The grayscale ablation (Table 7) demonstrating that rankings hold without color information is a thoughtful control. The profile-free ablation (Table 4, Table 8) provides evidence that profile conditioning improves data quality. The commitment to release data, code, and configurations upon acceptance supports reproducibility. The finding that multi-reference joint conditioning degrades performance is a useful empirical observation for the community.

major comments (4)
  1. §3.3 and §H.1: The candidate image pool for the Real-User dataset is generated using GPT-5 to produce profile-conditioned prompts, which are then rendered by Qwen-Image. Real users select 6–8 preferred images from this GPT-5-shaped candidate pool. When GPT-5-based VLM conditioning fusion subsequently achieves the highest Elo (1765) on this same data (Table 1), the benchmark may systematically favor methods whose aesthetic reasoning aligns with the LLM used to construct the preference space. This is a structural circularity that affects the headline claim. The paper's mitigations (multiple judges, 91% human agreement, Table 6 user study) address judge reliability but not the candidate-generation circularity. The authors should either (a) acknowledge this as a limitation and discuss its implications for the Elo comparison, or (b) provide evidence that the preference space is not biased — e
  2. §3.3, Synthetic-Agent preference selection: For synthetic agents, preferred images are selected via aesthetic scores rather than human judgment. The aesthetic scorer is not specified in the main text. Since the synthetic subset constitutes 719 of 1,369 test cases (52%), the choice of aesthetic model is load-bearing for half the benchmark. The authors should specify which aesthetic scorer is used and discuss whether it introduces systematic biases (e.g., favoring high-saturation or photorealistic images) that could confound the synthetic-agent evaluation results in Table 1.
  3. §3.1 and §K (Reference Table): The mapping from psychological traits to specific visual aesthetics (e.g., 'high Openness → abstract compositions,' 'low Openness → realistic, structured compositions') is stated as grounded in psychological literature, but the operationalization is entirely through LLM reasoning without independent empirical validation that these mappings hold for image generation preferences specifically. The psychological citations (e.g., [5, 19]) establish correlations between personality and art appreciation, not between personality and text-to-image prompt preferences. The paper should clarify the gap between the cited psychological findings and the specific visual-aesthetic mappings used in the pipeline, and note this as a limitation.
  4. Table 1, Real-User Elo column: The Elo ratings are computed only on the Real-User dataset (§4.2 states the arena evaluation uses 'our Real-User dataset'), yet Table 1 reports Elo for both Synthetic Agent and Real-User columns. It is unclear what the Synthetic Agent Elo values represent or whether they are computed differently. The authors should clarify whether Elo is computed for the synthetic subset and, if so, how (given that synthetic agents lack real human preference validation). If Elo is only computed for real users, the Synthetic Agent Elo column should be removed or relabeled.
minor comments (9)
  1. §3.2: The text states '134 valid responses' after quality control, but §H.1 states '124 valid questionnaires were retained.' The final count of 76 users is explained by 68 declining the second stage, but 134 − 68 = 66, not 76. Please reconcile these numbers.
  2. §3.4: The benchmark statistics state '1,369 test cases constructed from 1,876 images collected from 251 agents/users.' The breakdown (175 agents + 76 users = 251) is consistent, but the image count for the real-user subset (645 images for 650 test cases) seems low — please clarify how 650 test cases are derived from 645 images.
  3. Table 1: The no-preference baseline has different Elo values in Table 1 (1427) and Table 3 (1485) for the Real-User dataset. If these are from different evaluation runs, please note this; if not, please correct.
  4. §4.2: The confidence threshold for ties is stated as 0.70 in the supplementary (§B), but the axiom ledger lists it as a free parameter. Please state this threshold in the main text and briefly justify the choice.
  5. §3.2, Algorithm 1: The Jaccard similarity threshold τ_jac = 0.42 is mentioned in the supplementary (§J.4) but not in the main text. Including it in the algorithm description would improve self-containedness.
  6. Figure 2: The flowchart is difficult to parse — the relationship between 'Aesthetic Ranking,' 'Aesthetic Axes,' and 'Profile Keyword' boxes is unclear. Consider simplifying or adding a clearer caption.
  7. §4.3: The claim that 'aligning preferences for real users is easier than aligning those for synthetic agents' is supported by the observation that real users reflect 'prevailing preferences commonly seen on the internet.' This is plausible but speculative. Consider softening the language or providing supporting evidence.
  8. References: Several citations appear to be from 2025–2026 (e.g., [1] Qwen3-VL, [9] Gemini 2.5, [32] GPT-5). Please ensure all references are complete and accessible.
  9. §H.1: The text mentions '74 valid data entries' at one point, but the main text and abstract consistently use 76. Please correct.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The comments identify genuine issues that we will address in the revision. Below we respond point by point.

read point-by-point responses

  1. Referee: Candidate-generation circularity: GPT-5 generates profile-conditioned prompts for the candidate pool, and GPT-5-based VLM fusion subsequently achieves the highest Elo on the same data.

    Authors: We agree this is a legitimate structural concern. The candidate image pool for the Real-User dataset is shaped by GPT-5's profile-to-prompt reasoning, and GPT-5-based condition fusion later achieves the highest Elo on data derived from that pool. Our existing mitigations—multiple independent judges (GPT-5, Gemini 2.5 Pro, Qwen3-VL), 91% human agreement, and the Table 6 user study—address judge reliability but do not fully eliminate the candidate-generation circularity the referee identifies. We will add an explicit discussion of this limitation in the revised manuscript (Section O) and note its implications for the Elo comparison on the Real-User subset. We would also like to clarify several points that partially mitigate the concern: (1) Real users make genuine selections from the candidate pool—they choose 6–8 out of 20 images—so the final preference set reflects human judgment, not GPT-5's ranking. (2) The automatic metrics (CLS-R, DIS-R, LPIPS-R), which do not involve GPT-5, show the same ranking pattern on the Real-User data, with GPT-5 fusion achieving the highest CLS-R and DIS-R. (3) On the Synthetic Agent subset, where preference selection uses aesthetic scores rather than GPT-5-shaped candidates, GPT-5 fusion still achieves the best CLS-R and DIS-R. These observations suggest the ranking is not solely an artifact of candidate-generation circularity, though we acknowledge the concern cannot be fully dismissed without an experiment using a different LLM for candidate generation. We will note this as a direction for future work. revision: yes

  2. Referee: Aesthetic scorer for synthetic agents is unspecified; this is load-bearing for 52% of the benchmark.

    Authors: We agree this information should be specified in the main text. The aesthetic scorer used for synthetic-agent preference selection is LAION-AES v2 (the aesthetic prediction model based on CLIP ViT-L/14). We will add this specification to Section 3.3 and discuss potential biases. Regarding systematic biases: LAION-AES v2 is known to favor photorealistic, high-detail, and high-saturation images, which could confound the synthetic-agent evaluation. We note that the relative ranking of methods is consistent between the Synthetic Agent and Real-User subsets (e.g., GPT-5 fusion ranks first on both, Qwen-Image-Edit (1-Ref) outperforms (2-Ref) on both), which provides some evidence that the aesthetic scorer's biases do not fundamentally distort the method comparison. However, we will add a discussion of this potential confound as a limitation, particularly for absolute metric interpretation on the synthetic subset. revision: yes

  3. Referee: Psychological trait → visual aesthetic mappings lack independent empirical validation for text-to-image prompt preferences specifically.

    Authors: This is a fair observation. The cited psychological literature (e.g., Chamorro-Premuzic et al. [5], McCrae [19]) establishes correlations between personality traits—particularly Openness—and art appreciation or aesthetic preferences in traditional media. Our pipeline operationalizes these findings by mapping traits to specific visual-aesthetic parameters for text-to-image prompt generation, which is a different setting. The gap between the cited findings and our specific mappings has not been independently validated through controlled experiments. We will clarify this distinction in the revised Section 3.1 and add it to the Limitations section (Section O). We would note that our ablation in Table 8 provides partial evidence for the utility of these mappings: removing the Openness variable degrades data quality (lower Silhouette scores, higher entropy, lower human ratings), suggesting the psychological mappings contribute meaningfully to the pipeline. However, this does not constitute independent validation of the specific trait-to-visual correspondences (e.g., 'high Openness → abstract compositions'), and we will state this explicitly. revision: yes

  4. Referee: Table 1 reports Elo for both Synthetic Agent and Real-User columns; unclear what Synthetic Agent Elo represents given §4.2 states Elo uses the Real-User dataset.

    Authors: We appreciate the referee raising this point, as it indicates the table formatting may be ambiguous. To clarify: Elo is computed only on the Real-User dataset, as stated in §4.2. Table 1 does not contain a Synthetic Agent Elo column—the Synthetic Agent columns are CLS-T, LPIPS-R, CLS-R, and DIS-R only. The Elo column appears solely under the Real-User header. We believe the confusion may arise from the table's compact formatting. In the revised manuscript, we will improve the table formatting to make the column groupings clearer (e.g., with a visible separator or vertical rule between the Synthetic Agent and Real-User column groups) so that it is unambiguous that Elo is reported only for Real-User data. We will also add a footnote or caption note explicitly stating that Elo is computed exclusively on the Real-User subset. revision: yes

Circularity Check

0 steps flagged

GPT-5 generates the candidate image pool from which real users select preferences; GPT-5-based VLM fusion then wins on the resulting benchmark, creating a structural circularity in the headline Elo claim.

full rationale

The paper's headline result (GPT-5 VLM conditioning fusion achieves Elo 1765 on Real-User data) is partially circular because GPT-5 was used to construct the candidate image pool from which real users selected their preferred images (§H.1). The preference space available to users was shaped by GPT-5's profile-to-aesthetic reasoning, so when GPT-5-based fusion is evaluated on this data, it benefits from a preference space that aligns with its own reasoning patterns. This is a genuine structural circularity in the benchmark construction. However, it is not a pure self-definitional loop: real users made independent selections from the candidate pool, human annotators confirmed judge reliability (91% agreement), and a user study (Table 6) independently ranked GPT-5 first. The circularity is partial — the candidate generation constrains but does not fully determine user preferences. The synthetic agent pipeline (719 test cases) has a more severe circularity: LLM-generated profiles, LLM-generated prompts, and aesthetic-score-based preference selection all occur within the same automated pipeline, but this subset is not the source of the headline Elo claim. The paper does not acknowledge the candidate-generation circularity affecting the Real-User Elo results; its mitigations address judge reliability but not the structural overlap between candidate generation and the winning evaluation method. Score 5 reflects that the central claim has independent content (human selection, multiple judges, user study) but is partially forced by the construction of the preference space. No self-citation chain is involved; the circularity is architectural, not citation-based. No single step reduces exactly to its inputs by definition, so no individual step scores higher, but the aggregate effect on the headline claim is material.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The paper is a well-constructed benchmark contribution with a novel profile-inclusive design. The real-user data and human agreement studies provide genuine grounding. However, the synthetic agent pipeline (which constitutes 53% of test cases) has a circularity concern: LLMs generate profiles, prompts, and preference rankings, and the best-performing method uses similar LLM reasoning. The profile-to-aesthetic mappings are not independently validated for image generation specifically. The real-user sample is small and skewed. These issues do not invalidate the benchmark but warrant caution in interpreting synthetic-agent results and the superiority of LLM-based methods.

free parameters (5)
  • τ_conf (confidence threshold for agent acceptance) = 97 (implied by §J.3: score ≥ 97)
    Threshold for accepting synthetic agents based on soft-conflict penalty scoring; chosen by the authors.
  • τ_jac (Jaccard similarity threshold) = 0.42
    Diversity filter threshold for rejecting similar agents; stated in §3.2 and §J.4.
  • K (reference image count) = ≤5
    Number of preferred images sampled as reference set per test case; chosen by the authors.
  • M (preference tokens) = 32
    Number of learnable preference tokens in the cross-attention compressor; chosen by the authors (§E).
  • Tie confidence threshold = 0.70
    Threshold below which pairwise comparisons are classified as ties in the Elo rating system (§B).
axioms (4)
  • domain assumption Psychological traits (Openness, Schwartz values, EVT color associations, CMA affective dimensions) systematically map to specific visual aesthetic preferences.
    Invoked in §3.1 and operationalized throughout the pipeline (§3.3, §K). The paper cites psychological literature for individual mappings but does not empirically validate that these mappings hold specifically for image generation preferences.
  • ad hoc to paper LLMs can accurately reason about the relationship between user profiles and visual aesthetic dimensions to generate profile-aligned image prompts.
    The entire candidate image generation pipeline (§3.3) depends on LLM chain-of-thought reasoning producing prompts that genuinely reflect the user's implicit preferences. No independent validation that LLM-generated prompts match what a real user with that profile would prefer.
  • domain assumption Aesthetic scores (used for synthetic agent preference selection) are a valid proxy for human visual preference.
    Synthetic agent 'preferred' images are selected via aesthetic-score ranking (§3.3), assuming this correlates with what a real user with that profile would choose. Not independently validated.
  • domain assumption LLM-as-judge with persona conditioning provides a reliable proxy for human preference evaluation.
    The persona-aware Elo rating (§4.2) assumes that an LLM given a user profile can judge image preferences as that user would. Supported by ~91% agreement on a 200-case subset, but this is a limited validation.
invented entities (1)
  • Synthetic agents independent evidence
    purpose: Scalable generation of diverse user profiles and preference data to complement the small real-user dataset.
    The paper validates synthetic agents by showing consistency between synthetic and real-user results (§4.3) and by conducting a user study comparing profile-inclusive vs. profile-free pipelines (§4.5). However, the 'preferences' of synthetic agents are not validated against real humans with matching profiles.

pith-pipeline@v1.1.0-glm · 31471 in / 4503 out tokens · 173064 ms · 2026-07-08T05:35:11.012911+00:00 · methodology

0 comments
read the original abstract

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/

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ElasticTTT: Prior-Preserving Test-Time Tuning for Video Editing

    cs.CV 2026-07 conditional novelty 6.0

    Test-time tuning of video diffusion models collapses generation toward the source video; ElasticTTT counters this with noisy targets, contrastive source-prompt guidance, and asynchronous region-wise noise scheduling, ...

  2. ElasticTTT: Prior-Preserving Test-Time Tuning for Video Editing

    cs.CV 2026-07 conditional novelty 6.0

    A test-time tuning framework with three regularization techniques that preserves the generative prior of a video diffusion model during one-shot editing, achieving state-of-the-art results on the authors' benchmark.

Reference graph

Works this paper leans on

70 extracted references · 70 canonical work pages · cited by 1 Pith paper · 9 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Computer Science

    Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf2(3), 8 (2023)

    work page 2023

  4. [4]

    In: Social Stratification, Class, Race, and Gender in Sociological Perspective, Second Edition, pp

    Bourdieu, P.: Distinction: A social critique of the judgement of taste. In: Social Stratification, Class, Race, and Gender in Sociological Perspective, Second Edition, pp. 499–525. Routledge (2019)

    work page 2019

  5. [5]

    British Journal of Psychology100(3), 501–516 (2009)

    Chamorro-Premuzic, T., Reimers, S., Hsu, A., Ahmetoglu, G.: Who art thou? personality predictors of artistic preferences in a large uk sample: The importance of openness. British Journal of Psychology100(3), 501–516 (2009). https://doi.org/10.1348/000712608X366867

    work page doi:10.1348/000712608x366867 2009

  6. [6]

    In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining

    Chen, W., Huang, P., Xu, J., Guo, X., Guo, C., Sun, F., Li, C., Pfadler, A., Zhao, H., Zhao, B.: Pog: personalized outfit generation for fashion recommendation at alibaba ifashion. In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2662–2670 (2019)

    work page 2019

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, Z., Zhang, L., Weng, F., Pan, L., Lan, Z.: Tailored visions: Enhancing text-to-image generation with personalized prompt rewriting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7727–7736 (2024)

    work page 2024

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Chung, J., Hyun, S., Heo, J.P.: Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8795–8805 (June 2024)

    work page 2024

  9. [9]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Emu: Enhancing Image Generation Models Using Photogenic Needles in a Haystack

    Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023) College of AI, Tsinghua University | 14

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion

    Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  12. [12]

    Journal of Research in Personality37, 504–528 (04 2003)

    Gosling, S.D., Rentfrow, P.J., Swann, W.B.: A very brief measure of the big-five personality do- mains. Journal of Research in Personality37, 504–528 (04 2003). https://doi.org/10.1016/s0092- 6566(03)00046-1

    work page doi:10.1016/s0092- 2003

  13. [13]

    Acm transactions on interactive intelligent systems (tiis)5(4), 1–19 (2015)

    Harper, F.M., Konstan, J.A.: The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis)5(4), 1–19 (2015)

    work page 2015

  14. [14]

    In: Proceedings of the 2021 conference on empirical methods in natural language processing

    Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. In: Proceedings of the 2021 conference on empirical methods in natural language processing. pp. 7514–7528 (2021)

    work page 2021

  15. [15]

    Advances in Neural Information Processing Systems37, 79889– 79908 (2024)

    Jiang, D., Ku, M., Li, T., Ni, Y., Sun, S., Fan, R., Chen, W.: Genai arena: An open evaluation platform for generative models. Advances in Neural Information Processing Systems37, 79889– 79908 (2024)

    work page 2024

  16. [16]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kim, H., Ahn, S., Seo, Y.D.: Draw your mind: Personalized generation via condition-level modeling in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17171–17180 (2025)

    work page 2025

  17. [17]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 36652–36663. Curran Associates, Inc. (2023)

    work page 2023

  18. [18]

    Li, Y., Yang, S., Han, X., Wang, W., Dong, J., Lyu, Y., Xue, Z.: Instant preference alignment for text-to-image diffusion models (2025)

    work page 2025

  19. [19]

    Motivation and Emotion31(1), 5–11 (2007)

    McCrae, R.R.: Aesthetic chills as a universal marker of openness to experience. Motivation and Emotion31(1), 5–11 (2007)

    work page 2007

  20. [20]

    Mo, W., Ba, Y., Zhang, T., Bai, Y., Li, B.: Learning user preferences for image generation model (2025)

    work page 2025

  21. [21]

    arXiv preprint arXiv:2512.06020 (2025)

    Mo, W., Zhang, T., Bai, Y., Han, L., Ba, Y., Metaxas, D.N.: Prefgen: Multimodal preference learning for preference-conditioned image generation. arXiv preprint arXiv:2512.06020 (2025)

    work page arXiv 2025

  22. [22]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  23. [23]

    Proceedings of the National Academy of Sciences107(19), 8877–8882 (2010)

    Palmer, S.E., Schloss, K.B.: An ecological valence theory of human color prefer- ence. Proceedings of the National Academy of Sciences107(19), 8877–8882 (2010). https://doi.org/10.1073/pnas.0906172107

    work page doi:10.1073/pnas.0906172107 2010

  24. [24]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: Styleclip: Text-driven manipulation of stylegan imagery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2085–2094 (October 2021) College of AI, Tsinghua University | 15

    work page 2085

  25. [25]

    In: Yue, Y., Garg, A., Peng, N., Sha, F., Yu, R

    Peng, Y., Cui, Y., Tang, H., Qi, Z., Dong, R., Bai, J., han, c., Ge, Z., Zhang, X., Xia, S.T.: Dreambench++: A human-aligned benchmark for personalized image generation. In: Yue, Y., Garg, A., Peng, N., Sha, F., Yu, R. (eds.) International Conference on Representation Learning. vol. 2025, pp. 46010–46032 (2025)

    work page 2025

  26. [26]

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)

    work page 2022

  27. [27]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22500–22510 (2023)

    work page 2023

  28. [28]

    Journal of Personality and Social Psychology39, 1161–1178 (12 1980)

    Russell, J.: A circumplex model of affect. Journal of Personality and Social Psychology39, 1161–1178 (12 1980). https://doi.org/10.1037/h0077714

    work page doi:10.1037/h0077714 1980

  29. [29]

    In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gon- tijo Lopes, R., Karagol Ayan, B., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Informa...

    work page 2022

  30. [30]

    In: European Conference on Computer Vision

    Salehi, S., Shafiei, M., Yeo, T., Bachmann, R., Zamir, A.: Viper: Visual personalization of generative models via individual preference learning. In: European Conference on Computer Vision. pp. 391–406. Springer (2024)

    work page 2024

  31. [31]

    Online Readings in Psychology and Culture2(1) (12 2012)

    Schwartz, S.H.: An overview of the schwartz theory of basic values. Online Readings in Psychology and Culture2(1) (12 2012)

    work page 2012

  32. [32]

    OpenAI GPT-5 System Card

    Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al.: Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    In: Oxford Bibliographies in Psychology

    Soto, C.J., Jackson, J.J.: Five-factor model of personality. In: Oxford Bibliographies in Psychology. Oxford University Press (2020). https://doi.org/10.1093/obo/9780199828340-0120

    work page doi:10.1093/obo/9780199828340-0120 2020

  34. [34]

    In: European Conference on Computer Vision

    Von Rütte, D., Fedele, E., Thomm, J., Wolf, L.: Fabric: Personalizing diffusion models with iterative feedback. In: European Conference on Computer Vision. pp. 385–400. Springer (2024)

    work page 2024

  35. [35]

    Qwen-Image Technical Report

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    In: Proceedings of the 37th International Conference on Neural Information Processing Systems

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: learning and evaluating human preferences for text-to-image generation. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. pp. 15903–15935 (2023)

    work page 2023

  37. [37]

    Investigating Non-Transitivity in LLM-as-a-Judge

    Xu, Y., Ruis, L., Rocktäschel, T., Kirk, R.: Investigating non-transitivity in llm-as-a-judge. arXiv preprint arXiv:2502.14074 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    In: Proceedings of the ACM on Web Conference 2025

    Xu, Y., Wang, W., Zhang, Y., Tang, B., Yan, P., Feng, F., He, X.: Personalized image generation with large multimodal models. In: Proceedings of the ACM on Web Conference 2025. pp. 264–274 (2025)

    work page 2025

  39. [39]

    In: CVPR (2018) College of AI, Tsinghua University | 16

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) College of AI, Tsinghua University | 16

    work page 2018

  40. [40]

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

    Zhang, Z., Zhang, Q., Xing, W., Li, G., Zhao, L., Sun, J., Lan, Z., Luan, J., Huang, Y., Lin, H.: Artbank: Artistic style transfer with pre-trained diffusion model and implicit style prompt bank. Proceedings of the AAAI Conference on Artificial Intelligence38(7), 7396–7404 (Mar 2024). https://doi.org/10.1609/aaai.v38i7.28570 College of AI, Tsinghua Univer...

    work page doi:10.1609/aaai.v38i7.28570 2024

  41. [41]

    Major field of study Medicine Science

    work page

  42. [42]

    Laid-back

    Main role on campus “Laid-back” player Core member of a student organization and a “laid-back” player

    work page

  43. [43]

    Which style best describes your living space? Minimalist / functional style Cozy / healing style

    work page

  44. [44]

    Preferred travel type Urban exploration, exhibitions Natural scenery, historical sites, cultural tours

    work page

  45. [45]

    Preferred clothing style Simple style Sporty casual, simple style

    work page

  46. [46]

    Hobbies and interests Anime, comics, gaming, music, reading, interacting with pets Gaming, handicrafts, exploring nature, interacting with pets

    work page

  47. [47]

    Main content you consume or follow Fashion & beauty, pets, foodAnime commentary, game streaming, pets, food

    work page

  48. [48]

    Apps or platforms frequently used bilibili, Xiaohongshu, WeChat, Douyinbilibili, Xiaohongshu, Zhihu, WeChat, QQ, Douyin

    work page

  49. [49]

    Preferred thinking style Rating: 4 Rating: 4

    work page

  50. [50]

    I see myself as curious and open to new experiences

    "I see myself as curious and open to new experiences." Rating: 7 Rating: 3

    work page

  51. [51]

    I see myself as rigid in thinking and lacking creativity

    "I see myself as rigid in thinking and lacking creativity." Rating: 2 Rating: 4

    work page

  52. [52]

    Current level of pleasant mood Rating: 5 Rating: 4

    work page

  53. [53]

    Current level of emotional excitement Rating: 3 Rating: 0

    work page

  54. [54]

    Core personal pursuits (choose 3) Independence and creativity Novelty and excitement Hedonism: enjoyment Hedonism: enjoyment Security and stability

    work page

  55. [55]

    Frequency of using AI image generation in the past year Never used 1–3 times a year

    work page

  56. [56]

    minimalist and functional,

    Preferred color palette Urban minimalist tones Cool ocean tones retained. For each valid questionnaire, we invoked the GPT-5 API (hereafter GPT-API) to perform structured analysis of the text responses and the uploaded images. The GPT-API outputs consist of (1) a standardized user profile and (2) 20 diverse image-generation prompts automatically produced ...

    work page

  57. [57]

    Multi-select answers (Q3, Q5-Q9, Q16, Q18) recorded as entries of the form QX:OptionLabel” (e.g.,Q7:Photography)

    work page

  58. [58]

    Single-choice selections (e.g., field of study, gender, MBTI) encoded as Question=Option”

    work page

  59. [59]

    When a new candidate is generated, its signature is compared against those of the most recent 200 accepted agents

    Binned rating responses (Q11-Q15) expressed asLow/Med/High. When a new candidate is generated, its signature is compared against those of the most recent 200 accepted agents. TheJaccard similarity coefficientis used to quantify overlap; if similarity with any recent signature exceeds 0.42, the candidate is rejected and regenerated. This local-window compa...

    work page

  60. [60]

    Instruction Header. This section defines the model’s task: to infer a user’s latent image College of AI, Tsinghua University | 18 imagePsychological Explanationkey property (a)ENTPs prefer creative and abstract images with strong color contrast and abundant tension. High-contrast black and white.Strong light and shadow layers.Minimalist and abstract compo...

    work page

  61. [61]

    Questionnaire Item List. This component enumerates all referable user dimensions—discipline, gender, campus role, living style, travel and clothing preferences, hobbies, preferred content plat- forms, MBTI, TIPI, CMA, Schwartz values, AI usage frequency, and EVT palette choices. It also standardizes score ranges (Q11: 1–5; TIPI: 0–7; CMA: 0–9), ensuring c...

    work page

  62. [62]

    Additional Constraint: Diversity Enhancement for Prompt Generation. To prevent over-concentration of subject matter and encourage multimodal balance, the framework introduces a post-processing constraint during the generation of the 20 final prompts. It enforces:

    work page

  63. [63]

    Balanced category distribution:approximately 25–30% portraits, 20% landscapes, 15% urban scenes, 15% still-life, 10% abstract visuals, and 10–15% creative hybrids

    work page

  64. [64]

    Rotated focal subjects:at least one prompt per major category (nature, urban, still-life, abstract, cultural/narrative)

    work page

  65. [65]

    Limited repetition:no more than two prompts may share the same dominant subject or setting

    work page

  66. [66]

    Encouraged non-human focus:inclusion of atmospheric or object-centered scenes without explicit human subjects

    work page

  67. [67]

    Maintained stylistic coherence:all prompts must adhere to the aesthetic profile inferred from questionnaire and image analysis (color palette, mood, realism, and compositional tone)

    work page

  68. [68]

    Campus Roles: .../ Living Space Style: .../ Color Preferences:

    User Profile Slot. This slot embeds the individual profile text parsed from the agent dataset, such as “Campus Roles: .../ Living Space Style: .../ Color Preferences: ....” It serves as a bridge between structured agent data and natural language reasoning, allowing large language models to access user context

    work page

  69. [69]

    Reference Table (Theory-to-Visual Map). The reference table operationalizes psychological, behavioral, and demographic frameworks into visual reasoning rules that link human traits and College of AI, Tsinghua University | 20 contexts to aesthetic outputs. These mappings transform high-level constructs—personality, values, affect, lifestyle, and sensory pr...

    work page

  70. [70]

    User-preferred

    Output Schema (Generation Format Reference)˙The output schema defines a reproducible data structure encompassing key psychological metrics, inferred traits (Openness, CMA, EVT, Schwartz), contextual synthesis of Q1–Q11, one-sentence style summary, actionable generation parameters (style tags, lighting, composition, constraints), scenario recommendations, ...

    work page