Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Generate, Not Recommend: Personalized Multimodal Content Generation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A generative recommender that creates the image of a user's next item, trained with future-item rewards, beats retrieval baselines on personal relevance.

desk verdict New task, but the headline future-relevance results are optimized rather than demonstrated; the paper's own text confirms the reward includes future items. read the letter →

arxiv 2506.01704 v1 pith:Y7SCUQAD submitted 2025-06-02 cs.AI cs.CL

Jiongnan Liu , Zhicheng Dou , Ning Hu , Chenyan Xiong This is my paper
classification cs.AIcs.CL
keywords personalizedcontentgenerationmultimodallargemodelsonlinereinforcementlearningGRPOrecommendersystemssequentialrecommendationimage
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish a new paradigm for personalization: instead of recommending items from an existing catalog, a model should directly generate the image content a user will like next. It claims that an any-to-any large multimodal model, fine-tuned first by supervised learning on user interaction histories and then by online reinforcement learning with rewards tied to personal relevance, produces movie posters and video covers that are more relevant to users' past and future preferences than content retrieved by a state-of-the-art recommender. The experiments compare the generated images with retrieved ones on two standard recommendation datasets using semantic and perceptual similarity metrics, and a small user study also prefers the generated images. If the claim holds, personalized systems could move from filtering existing content to creating novel content for each user.

What carries the argument

The central mechanism is the any-to-any large multimodal model Janus-Pro-1B used as a generation policy, trained in two stages. In the supervised fine-tuning stage, the model is trained to produce the image tokens of a user's k-th item conditioned on the previous k-1 items' text and images, using randomly masked and swapped histories to expand sparse interaction data. In the online RL stage, Group Relative Policy Optimization is applied: for each history the policy samples G images, computes a group-normalized advantage from a reward that sums scaled similarity scores — CLIP text and image matching (CTS/CIS), DINO image similarity (DIS), LPIPS/MS-SSIM/SSIM perceptual metrics, a profile-based CLIP score (PCS), and NIMA aesthetics — measured against historical items, the golden next item, and potential future items, with KL regularization to the SFT policy. This machinery lets the model explore content beyond observed interactions, which the paper argues is necessary because a matching item may be absent from the logged history.

What would settle it

Train the RL stage with the actual next and future items removed from the reward (only history-derived relevance and aesthetics), then evaluate on a held-out test set; if the future-relevance metrics (CTS/CIS/DIS to the future items) no longer beat the HLLM retrieval baseline, the claim that the model anticipates future interests would be refuted.

Watch

Extended reading notes

Core claim

This paper claims that a generative model can take over the personalization step that recommender systems currently perform by retrieval: given only a user's text-and-image interaction history, an any-to-any large multimodal model can directly generate the visual content of that user's next item, with no candidate or reference image supplied. The training pipeline first imitates the next item's image tokens from the previous five interactions, then applies GRPO online reinforcement learning, where each sampled image receives a reward that adds normalized semantic and perceptual relevance scores against the user's historical items, the actual next item, and potential future items, plus an aesthetic score. On two recommendation benchmarks, the paper reports that the best of four generated images from the fully trained model beats a state-of-the-art retrieval recommender (HLLM) on most of these relevance metrics, for all three comparison sets, and that RL improves over SFT on metrics where SFT lags. The authors also report that the RL model can fall into reward hacking by repeatedly generating a popular poster (The Godfather), which they attribute to CLIP's familiarity with popular images, and therefore select their final checkpoint before the reward standard deviation collapses.

Load-bearing premise

The RL reward uses the actual next item and potential future items that the user will interact with, but at deployment those future items are unknown, so the model is trained on the very targets it is later evaluated against.

Editorial extensions

If this is right

  • Recommendation can become content synthesis: a system could generate a novel poster, cover, or artwork for each user without needing a candidate from a catalog.
  • Online reinforcement learning with relevance rewards measurably improves generation quality beyond supervised fine-tuning, especially on metrics where SFT lags, suggesting exploration helps under sparse interaction logs.
  • The reported gains on potential-future-item metrics indicate the model captures longer-term interest drift, not just repetition of historical patterns.
  • The approach transfers across domains — movie posters and micro-video covers — with the same training recipe and backbone.
  • The observed reward hacking implies that relevance-from-CLIP rewards carry a popularity bias, so deployments must monitor reward variance and stop before collapse.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the oracle access to future items is removed at deployment, the future-relevance advantage likely shrinks; a natural test is to train with predicted next-item embeddings as pseudo-targets and evaluate on a strict temporal split.
  • The reward-hacking episode suggests the reliance on CLIP/DINO scorers creates a popularity bias in any learned similarity reward; replacing the relevance reward with a human-preference signal would test whether the method's gains transfer beyond the metric suite.
  • A direct extension is to check whether generated images retain their relevance edge when judged by a different scorer or by fresh human annotators; if the gains vanish, the method is overfitting to the fixed metric set.
  • The paper's success at using retrieval datasets as generation supervision hints that larger interaction logs collected for recommendation could fuel generative personalization, an avenue the authors do not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a new paradigm, 'Generate, Not Recommend', in which an any-to-any large multimodal model (Janus-Pro-1B) is fine-tuned in two stages to generate personalized images (movie posters and micro-video covers) directly from a user's multimodal interaction history, without relying on a candidate or reference item. The first stage is supervised fine-tuning on the next-item prediction task; the second stage applies GRPO with rewards that combine aesthetics and the personal relevance of generated images to historical, actual-next, and potential-future items. The authors evaluate on MovieLens and PixelRec, compare against the HLLM retrieval-based recommender, and report a set of semantic, perceptual, and aesthetic metrics as well as a small user study. The paper's central claim is that the generated images are more personally relevant to users' past and potential future preferences than content retrieved by state-of-the-art recommendation models.

Significance. If validated, the proposed paradigm would open a meaningful new direction beyond retrieval-based recommendation, and the two-stage SFT+RL training recipe on LMMs is a plausible way to approach it. The paper provides a reasonable starting testbed built from two public recommendation datasets, and it includes a candid analysis of reward hacking. However, the current evidence does not support the headline claim: the evaluation metrics are also used as reward signals in GRPO, the reward includes oracle access to actual and future items, and the reported comparisons are not statistically grounded. These issues are load-bearing for the central claim, and the user study does not cover the future-relevance dimension or compare against HLLM.

major comments (4)
  1. [§2.2, §3] The GRPO reward is computed from the same metrics (CTS, CIS, DIS, LPIPS, SSIM, MS-SSIM, PCS, NIMA) that are later reported as evaluation results in Table 1; Section 3 explicitly states that 'All the above metrics also serve as reward signals in our GRPO algorithm'. Since HLLM is not optimized with these rewards, the comparison measures how well the proposed model optimizes its own training objective rather than an independent capability. To support the claim that the generated images are more relevant to potential future preferences than content retrieved by a state-of-the-art recommender, the authors must report at least one evaluation metric that was not part of the reward, or train a baseline under the same reward scheme; without this, the central claim is not established.
  2. [§2.2] The reward signal includes the 'actual next' and 'potential future' items (Huk and Huk+1:k+p). At deployment these items are unknown, so the model is trained with oracle access to future interactions. The high scores in the 'Relevance to Potential Future Items' rows of Table 1 may therefore reflect training-time supervision from those very items, not an emergent ability to anticipate unseen future interests. The paper should ablate the future-reward terms (or retrain without them) to show that the future-relevance scores are not simply a consequence of optimizing for the future labels.
  3. [Table 1, Appendix C] Table 1 reports the best score among the top 4 generated or retrieved images, and Appendix C states that the final RL model is selected based on the highest average reward 'before any sign of reward hacking'. Neither the best-of-4 selection nor the checkpoint selection is accompanied by variance or significance measures across seeds or test samples. This reporting convention inflates performance and makes the differences in Table 1 difficult to interpret; please report means with standard deviations (and ideally confidence intervals or paired tests).
  4. [§4.2, Table 2] The user study compares only Janus+SFT against Janus+SFT+RL on 30 MovieLens cases, asking annotators which posters are high-quality. It does not compare against HLLM, and it does not measure relevance to future preferences. Consequently, the user study cannot corroborate the headline claim that the generated content is more personally relevant to potential future interests than content retrieved by state-of-the-art recommendation models.
minor comments (4)
  1. [§1] The sentence 'this retrieval-based paradigm may fails to fully fulfill user needs' contains a subject-verb agreement error ('may fails').
  2. [Table 1] Table 1 is difficult to read because the model names and dataset names are packed into a single header row; using separate column groups for each dataset and model would improve clarity.
  3. [§3] The PCS metric uses a Qwen-2.5-VL 7B model for user profiles, while the rest of the pipeline uses Janus-Pro-1B; please clarify whether the same Qwen model is applied to the HLLM baseline and the proposed model, and whether this introduces an additional pretrained model that is not fine-tuned.
  4. [§4.1] Observation 2 states that GRPO 'significantly boosts performance' on CIS, but no statistical significance test is reported; please either add significance tests or soften the claim.

Circularity Check

3 steps flagged · score 6.0 of 10

Future-relevance scores are the GRPO reward, not a prediction: Section 3 says all metrics double as rewards, Section 2.2 rewards the actual next and future items, and Appendix C selects the final model on the same reward.

  1. fitted input called prediction [Section 3 (Experimental Setup, 'Datasets & Metrics'), final sentence of metric definition; tied to L_GRPO in Section 2.2]
    "All the above metrics also serve as reward signals in our GRPO algorithm (Section 2.2)."

    This sentence makes every number in Table 1—CTS, CIS, DIS, LPIPS, SSIM, MS-SSIM, PCS, and NIMA—part of the training reward. The reported 'Relevance to Potential Future Items' rows are therefore not predictions from a fixed policy; they are values of the exact composite objective the policy was trained to maximize. The paper frames these scores as evidence that the model generalizes to future interests, but the evaluation is the optimized loss, which is the classic fitted-input-called-prediction pattern.

  2. fitted input called prediction [Section 2.2, paragraph after the GRPO objective]
    "Specifically, we incorporate the personal relevance of generated images to the historical interacted, the actual next, and the potential future contents as GRPO’s reward signals."

    The reward oracle receives the golden next item and potential future items. At deployment those items are unknown, so the model is not being evaluated on an anticipatory capability; it is being evaluated on an objective computed from the very future contents it is claimed to predict. Even with a train/test split, training-time use of future labels means the future-relevance metric measures reward optimization, not out-of-sample prediction of unseen preferences.

1 more flagged steps
  1. fitted input called prediction [Appendix C, 'Online Reinforcement Learning' bullet]
    "The final model is selected based on the highest average reward before any sign of reward hacking (e.g., sudden reward standard deviation drops)."

    Checkpoint selection is itself performed on the same reward that includes the future-item relevance metrics and all Table 1 metrics, and Table 1 reports the best of four generated images. The reported scores are therefore an in-sample selection statistic on the training objective, not an independent evaluation. The paper's own reward-hacking analysis concedes the metric prefers popular posters, so the selected numbers reflect shortcut exploitation avoided by early stopping.

full rationale

The SFT stage is a legitimate supervised training step: it uses user histories to generate the next item image, which is a self-contained objective. The user study is also an independent (though narrow) comparison: it only contrasts Janus+SFT against Janus+SFT+RL on 30 MovieLens cases and does not measure future relevance or compare against HLLM. The main circularity is concentrated in the RL stage and the headline evaluation. Section 3 explicitly states that every reported relevance and aesthetics metric is also a GRPO reward, and Section 2.2 states that the reward includes the actual next item and potential future items. Therefore the 'Relevance to Golden Next Item' and 'Relevance to Potential Future Items' rows of Table 1 report quantities the policy was directly optimized to maximize; they are fitted values of the training reward rather than independent demonstrations that the model anticipates unseen future tastes. Appendix C compounds this by selecting the final checkpoint on the same average reward, and Table 1 reports the best of four generated images, making the results an in-sample, best-of selection statistic. The train/test split prevents standard label leakage, but it does not establish the deployment capability claimed, because future items are not available at inference. The limitation passage concedes that the fine-tuned model 'tends to generate images that closely resemble those seen during training,' reinforcing that the future-relevance 'prediction' is a reward fit, not demonstrated generalization. No load-bearing self-citation was found; the cited baselines, metrics, and backbone models are external. Overall: partial circularity, score 6.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several choices that are not independently justified: the use of future items as reward signals, the choice of similarity metrics as both reward and evaluation, and the sliding-window task framing. No new entities are introduced; the method uses existing models and algorithms.

free parameters (7)
  • KL penalty coefficient beta = 0.04
    Set by hand for GRPO; controls policy drift from the reference model.
  • Rollout size G = 8
    Number of candidate images sampled per prompt during RL; larger values would improve exploration but are limited by compute.
  • SFT learning rate = 4e-5
    Chosen for supervised fine-tuning of Janus-Pro-1B.
  • RL learning rate = 1e-6
    Chosen for GRPO; small to avoid destabilizing the policy.
  • Sliding window sizes = k=6 (5 history, 1 next), p=3 (future)
    Task framing: five past items condition the next item, with three future items used for evaluation and as RL rewards. This window size is not justified.
  • Top-4 selection for evaluation = 4
    For each user, the best-scoring image among 4 generated or retrieved items is reported, inflating metric values.
  • Reward scaling and weighting = [0,1] each, unweighted sum
    All reward metrics are scaled and summed without tuned weights; the equal weighting is an ad hoc choice.
assumptions (5)
  • domain assumption Janus-Pro-1B can be fine-tuned to generate semantically meaningful images from multimodal histories.
    The authors show in Appendix B that vanilla Janus generates meaningless color blocks, and they assume SFT unlocks this ability; this is an empirical premise not proven in the paper.
  • domain assumption CLIP, DINO, LPIPS, SSIM and similar similarity scores reflect personal relevance of generated content.
    These metrics are used as both rewards and evaluation, but the paper itself observes they favor popular images (reward hacking in Section 4.2), calling this assumption into question.
  • ad hoc to paper The sliding-window framing (5 history, 1 next, 3 future) is a valid proxy for the personalized generation task.
    No justification is given for these window sizes; they are inherited from sequential recommendation conventions.
  • ad hoc to paper Future items are available for computing rewards during RL training.
    This is the key leak: the model is optimized to match ground-truth next and future items, which are not available at deployment, and this assumption directly enables the reported future-relevance results.
  • standard math GRPO (Group Relative Policy Optimization) works for this LMM generation setting.
    Borrowed from Shao et al. 2024 as an accepted algorithm; the paper does not validate its applicability to multimodal generation beyond the observed results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generate, Not Recommend: Personalized Multimodal Content Generation." pith.science (2026). https://pith.science/paper/Y7SCUQAD

@misc{pith2026250601704,
  author       = {Pith},
  title        = {Pith review of: Generate, Not Recommend: Personalized Multimodal Content Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y7SCUQAD}},
  note         = {Machine review of arXiv:2506.01704}
}
read the original abstract

To address the challenge of information overload from massive web contents, recommender systems are widely applied to retrieve and present personalized results for users. However, recommendation tasks are inherently constrained to filtering existing items and lack the ability to generate novel concepts, limiting their capacity to fully satisfy user demands and preferences. In this paper, we propose a new paradigm that goes beyond content filtering and selecting: directly generating personalized items in a multimodal form, such as images, tailored to individual users. To accomplish this, we leverage any-to-any Large Multimodal Models (LMMs) and train them in both supervised fine-tuning and online reinforcement learning strategy to equip them with the ability to yield tailored next items for users. Experiments on two benchmark datasets and user study confirm the efficacy of the proposed method. Notably, the generated images not only align well with users' historical preferences but also exhibit relevance to their potential future interests.

Figures

Figures reproduced from arXiv: 2506.01704 by the authors.

Figure 1
Figure 1. Personalized content provision paradigms. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Generation cases in MovieLens dataset. 0 100 200 300 400 500 600 700 800 Training Steps 0.1 0.2 0.3 0.4 0.5 MovieLens Reward STD 0.75 0.80 0.85 0.90 0.95 1.00 1.05 PixelRec Reward STD Reward STD Curves MovieLens Reward std Reward std curves PixelRec Rew ard std [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Generation cases by Vanilla Janus-Pro-1B [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: Instruction template of Janus-Pro-1B to generate personalized movie posters on MovieLens dataset based [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Instruction template of Janus-Pro-1B to generate personalized micro-video cover on PixelRec dataset [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Temporal Interest-Driven Multimodal Personalized Content Generation

    cs.IR 2025-09 reject novelty 2.0 of 10

    TIMGen is an unvalidated architecture proposal that combines Transformer temporal interest modeling, attention-based multimodal fusion, and VAE generation; the paper reports no experiments and no evaluation.

Reference graph

Works this paper leans on

22 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Miriam Arnold, Mascha Goldschmitt, and Thomas Rigotti. 2023. Dealing with information overload: a comprehensive review. Frontiers in psychology, 14:1122200

    work page 2023

  4. [4]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Ming - Hsuan Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. https://doi.org/10....

    arXiv doi:10.48550/arxiv.2502.13923 2025

  5. [5]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. https://doi.org/10.48550/ARXIV.2409.12740 HLLM: enhancing sequential recommendations via hierarchical large language models for item and user modeling . CoRR, abs/2409.12740

    arXiv doi:10.48550/arxiv.2409.12740 2024

  6. [6]

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. 2025. https://doi.org/10.48550/ARXIV.2501.17811 Janus-pro: Unified multimodal understanding and generation with data and model scaling . CoRR, abs/2501.17811

    arXiv doi:10.48550/arxiv.2501.17811 2025

  7. [7]

    Yu Cheng, Yunzhu Pan, Jiaqi Zhang, Yongxin Ni, Aixin Sun, and Fajie Yuan. 2024. https://doi.org/10.1137/1.9781611978032.49 An image dataset for benchmarking recommender systems with raw pixels . In Proceedings of the 2024 SIAM International Conference on Data Mining, SDM 2024, Houston, TX, USA, April 18-20, 2024 , pages 418--426. SIAM

    work page doi:10.1137/1.9781611978032.49 2024

  8. [8]

    Hossein Talebi Esfandarani and Peyman Milanfar. 2018. https://doi.org/10.1109/TIP.2018.2831899 NIMA: neural image assessment . IEEE Trans. Image Process. , 27(8):3998--4011

    work page arXiv 2018

Show all 22 references
  1. [9]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2016. https://doi.org/10.1145/2827872 The movielens datasets: History and context . ACM Trans. Interact. Intell. Syst. , 5(4):19:1--19:19

    2016 doi

  2. [10]

    Wang - Cheng Kang and Julian J. McAuley. 2018. https://doi.org/10.1109/ICDM.2018.00035 Self-attentive sequential recommendation . In IEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018 , pages 197--206. IEEE Computer Society

    2018

  3. [11]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023 a . http://papers.nips.cc/paper\_files/paper/2023/hash/6dcf277ea32ce3288914faf369fe6de0-Abstract-Conference.html Visual instruction tuning . In Advances in Neural Information Processing Systems 36: Annual Conference...

    2023

  4. [12]

    Zhenghao Liu, Sen Mei, Chenyan Xiong, Xiaohua Li, Shi Yu, Zhiyuan Liu, Yu Gu, and Ge Yu. 2023 b . https://doi.org/10.1145/3583780.3615077 Text matching improves sequential recommendation by reducing popularity biases . In Proceedings of the 32nd ACM International Conference on...

    2023

  5. [13]

    Maxime Oquab, Timoth \' e e Darcet, Th \' e o Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El - Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po - Yao Huang, Shang - Wen Li, Ishan Misr...

    2024

  6. [14]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. http://proceedings.mlr.press/v139/radford21a.html Learning transferable visual models...

    2021

  7. [15]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. https://doi.org/10.48550/ARXIV.2402.03300 Deepseekmath: Pushing the limits of mathematical reasoning in open language models . CoRR, abs/2402.03300

    2024 doi arXiv

  8. [16]

    Xiaoteng Shen, Rui Zhang, Xiaoyan Zhao, Jieming Zhu, and Xi Xiao. 2024. https://doi.org/10.1145/3589334.3645633 PMG : Personalized multimodal generation with large language models . In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024 , pages ...

    2024

  9. [17]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024 a . https://doi.org/10.48550/ARXIV.2409.1...

    2024 doi arXiv

  10. [18]

    Xianquan Wang, Likang Wu, Shukang Yin, Zhi Li, Yanjiang Chen, Hufeng Hufeng, Yu Su, and Qi Liu. 2024 b . https://aclanthology.org/2024.emnlp-main.1187 I-AM-G: interest augmented multimodal generator for item personalization . In Proceedings of the 2024 Conference on Empirical ...

    2024

  11. [19]

    Zhou Wang, Eero P Simoncelli, and Alan C Bovik. 2003. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, pages 1398--1402. Ieee

    2003

  12. [20]

    Yiyan Xu, Wenjie Wang, Yang Zhang, Biao Tang, Peng Yan, Fuli Feng, and Xiangnan He. 2025 a . https://doi.org/10.1145/3696410.3714843 Personalized image generation with large multimodal models . In Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia,...

    2025

  13. [21]

    Yiyan Xu, Wuqiang Zheng, Wenjie Wang, Fengbin Zhu, Xinting Hu, Yang Zhang, Fuli Feng, and Tat-Seng Chua. 2025 b . Drc: Enhancing personalized image generation via disentangled representation composition. arXiv preprint arXiv:2504.17349

    2025 arXiv

  14. [22]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. 2018. https://doi.org/10.1109/CVPR.2018.00068 The unreasonable effectiveness of deep features as a perceptual metric . In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, ...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.