Pith. sign in

REVIEW 3 major objections 5 minor 59 references

CTR-Driven Advertising Image Generation with Multimodal Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read A multimodal LLM pipeline can generate advertising images optimized for clicks, and a live test shows a 7.4% relative CTR gain over a pre-trained baseline.

desk verdict A commercially credible CTR-driven image generation system with a real online A/B test, but the reward model's out-of-distribution reliability is under-proven. read the letter →

arxiv 2502.06823 v1 pith:FM6PVKGA submitted 2025-02-05 cs.LG cs.CVcs.GRcs.IR

classification cs.LGcs.CVcs.GRcs.IR
keywords CTR-drivenadvertisingimagegenerationmultimodallargelanguagemodelsrewardmodelpreferenceoptimizatione-commerceclick-throughratepredictionproduct-backgroundmatchingdiffusion
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

Advertising images are usually generated to look good, not to get clicked. This paper argues that click-through rate should be the primary objective, and it builds a pipeline that optimizes for clicks while keeping the background relevant to the product. The method, CAIG, pre-trains a multimodal large language model on e-commerce data, trains a reward model that compares pairs of product images and predicts which one users will click more, then fine-tunes the prompt model with direct preference optimization plus a product-centric term that penalizes product-background mismatches. The authors report that this approach outperforms existing open- and closed-source baselines on pairwise click prediction, and that it improves online CTR by 7.4% relative to a pre-trained MLLM baseline across all tested product categories, with a separate A/B test gaining 2% over 60 million impressions.

What carries the argument

The load-bearing machinery is the two-branch reward model plus the product-centric preference loss. To compare two candidate advertising images, the model concatenates them into one image, adds a text prompt built from the product's title, category, tags, and other attributes, and runs the whole input through a multimodal LLM. The final token's hidden state feeds two heads: a classification head that decides which side has higher CTR, and a regression head that predicts absolute CTR values. This pairwise formulation avoids the need to calibrate CTR across product categories, since the model only ever judges images of the same product. The preference-optimization stage then fine-tunes the prompt model with the standard DPO loss, and adds a PCPO loss in which preference pairs are constructed by masking the product image or swapping in another product's text, forcing the generator to condition on the product rather than on the background alone.

What would settle it

Show images generated by the fine-tuned prompt model online for long enough to collect per-image CTR at the same confidence thresholds the paper uses for its RM test set (at least 1,000 impressions per image and at least a 5% relative CTR difference within pairs), then compute the reward model's pair accuracy on those pairs. If accuracy drops to near chance on the subset with CTR differences below 5%, the preference labels used for DPO cannot be trusted and the reported CTR gains would need another explanation.

Watch

Extended reading notes

Core claim

The central claim is that a multimodal large language model can serve as both the judge and the optimizer in a CTR-driven advertising image pipeline. The paper builds a two-branch reward model from a base MLLM pre-trained with e-commerce knowledge: a classification head compares two images of the same product and predicts which will get more clicks, while a regression head predicts absolute CTR values as an auxiliary signal. This reward model labels pairs of generated images, and those labels drive DPO fine-tuning of a prompt model that writes background descriptions for a diffusion-based generator. To keep the backgrounds relevant, the paper adds Product-Centric Preference Optimization, which corrupts the product image or text in half of each preference pair so the model learns to tie the background to the product's own attributes. With this setup it reports 58.6% pair accuracy on commercial data and 56.2% on public data, and online gains of 7.4% relative CTR improvement over the pre-trained baseline across all categories plus a 2% gain in a 60-million-impression A/B test.

Load-bearing premise

Everything depends on the reward model's judgments: it is trained on historical image pairs filtered to have large CTR differences, but in the RL loop it must provide reliable relative judgments for freshly generated pairs whose true CTR differences may be much smaller, so if it only separates easy pairs the DPO labels are mostly noise.

Editorial extensions

If this is right

  • Ad creative pipelines can be driven by an explicit click objective, so generation quality is measured by user clicks rather than by human aesthetic ratings.
  • The pairwise reward formulation avoids per-category CTR calibration, making the same reward model usable across heterogeneous product categories.
  • Adding relevance-preserving negative pairs through PCPO prevents the common reinforcement-learning failure where a background becomes visually attractive but semantically wrong.
  • Because only the prompt model is fine-tuned, the approach keeps the expensive diffusion generator frozen and still changes the generated content substantially.
  • A single reward model trained on historical click pairs can provide training signal for an entire advertising image pipeline, replacing labor-intensive aesthetic annotation.

Reading between the lines

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

  • A natural check the paper does not run is whether the reward model's accuracy on the filtered test set (1,000 impressions and 5% CTR difference) transfers to the generated images the prompt model actually produces, whose true CTR gaps are likely smaller; without that check, part of the DPO signal may be noise.
  • The 7.4% and 2% online gains are measured against a pre-trained MLLM baseline and the platform's existing traffic, not against the best already-deployed production system, so a strong incumbent optimizer might see smaller margins.
  • The same judge-and-optimize loop could transfer to other generative settings where user behavior is the reward, such as video thumbnails or recommendation-card images, provided historical pairwise behavior data exists.
  • Because the reward model learns from historical clicks, it inherits existing click biases such as position and brand effects, and the paper's own limitation note about aggregated user preferences points toward personalized RLHF as the next step.
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

3 major / 5 minor

Summary. The paper proposes CAIG, a pipeline for CTR-driven advertising image generation. It first pre-trains an MLLM (LLaVA-v1.6-7B) on e-commerce tasks to inject domain knowledge, then trains a two-branch reward model on pairwise CTR comparisons from historical impression/click logs, using a classification head for relative preference and a regression head for absolute CTR. In the final stage, the prompt model is fine-tuned with Direct Preference Optimization (DPO) and a proposed Product-Centric Preference Optimization (PCPO) term, where the reward model labels pairs of images generated by Stable Diffusion and ControlNet. Offline experiments report Reward Model pair accuracy of 58.6% on commercial data and 56.2% on public data, and online experiments report a 7.4% relative CTR improvement over the pre-trained MLLM baseline, plus a separate A/B test with a 2% CTR gain over 60 million impressions. The paper also ablates the reward model components, compares PCPO with standard DPO on a match-rate metric, and releases code and pre-trained models.

Significance. If the reported results hold, CAIG is a meaningful industrial contribution: it is, to my knowledge, the first work to use MLLMs as the core of a CTR-driven advertising-image generation pipeline, and the online A/B evidence is a genuine external anchor for the central claim. The paper also ships reproducible assets (code and pre-trained models) and uses real user click labels rather than synthetic preferences, which is a strength. The PCPO idea of enforcing product-centric consistency during preference optimization is useful and appears to mitigate a real failure mode of DPO. However, the significance is tempered by three issues: the offline reward-model accuracy is only modestly above chance even on easy test pairs; no statistical significance testing is reported for the online CTR differences; and the reward model is never validated on the generated-image distribution on which it is used to label DPO pairs. These gaps make the paper's strongest claim, that the generator is genuinely optimized for click preference, currently under-supported.

major comments (3)
  1. [§4.2.2, §4.1, Algorithm 1, Eq. (10)] The reward model is trained and evaluated only on historical advertising-image pairs filtered to require at least 1,000 impressions and a relative CTR difference above 5% in the test set (§4.1). In Algorithm 1, the same RM labels pairs of freshly generated images produced by Stable Diffusion and ControlNet, whose visual distribution differs from historical creatives and whose CTR differences may be much smaller. The paper never evaluates RM accuracy on this deployment distribution. Given that the offline Pair Accuracy is only 58.6% on commercial data and 56.2% on public data even on the easy, high-difference test pairs, the DPO signal in Eq. (10) could be mostly noise when applied to generated pairs. This is a load-bearing issue: the paper's central novelty is that the generator is optimized by genuine click preferences, and that attribution depends on the RM transmitting reliable preference labels. Please add a held-out evaluation of the RM on generated image pairs labeled by real user clicks (or by human raters as a proxy), ideally stratified by the size of the CTR difference. Without this, the causal claim that the observed online gains come from CTR-driven optimization is not established.
  2. [§4.4, Table 2] The headline online results are reported as point estimates without any uncertainty quantification. Table 2 gives relative CTR improvements of 7.4% across all categories and 3.2% to 9.5% in individual categories, and the A/B test reports a 2% CTR gain over 60 million impressions, but no confidence intervals, p-values, or Bayesian credible intervals are provided. The text says the one-week experiment 'validates the reliability and statistical significance of the CTR results,' yet no significance test is shown. The claim of 'significantly improving online CTRs' in the abstract and conclusion is therefore not formally supported. Please report confidence intervals and appropriate significance tests, and clarify the relationship between the 10-million-impression experiment and the 60-million-impression A/B test (are they separate deployments or cumulative counts?).
  3. [§4.3.2, Figure 4, Table 2] The advantage of PCPO over standard DPO is assessed primarily through the match rate, which is based on subjective annotations by 'experienced advertising professionals' with no reported inter-annotator agreement, annotation count, or sample construction details beyond 'randomly selecting 1,000 products.' The online comparison in Table 2 also includes a DPO row, but the online pipeline additionally involves professional human pre-filtering of displayed images (§4.4), a multi-armed bandit allocation policy, and the pre-trained MLLM baseline differences. These components are not held fixed or quantified across the compared methods, so the online gain cannot be cleanly attributed to the PCPO objective. Please provide inter-annotator agreement statistics for the match-rate evaluation and, ideally, an online ablation that isolates the PCPO contribution while keeping the human filtering and display policy fixed.
minor comments (5)
  1. [§3.4, Eq. (8)] Equation (8) is written as an update for x_t in terms of x_{t+1}, which is not the standard DDIM update and is difficult to verify; please align the notation with the cited DDIM formulation or explicitly state the denoising direction.
  2. [§3.4, Eq. (11) and surrounding text] The notation for the mismatched product information is inconsistent: the text uses 'd𝐼𝑜', the equation uses 'R𝐼𝑜' (or a similar symbol), and the pair is written as '(𝑦+, d𝐼𝑜, 𝐶)' rather than as a complete preference pair with a negative example. Please unify the notation and define the mismatched input precisely.
  3. [§3.3, Eq. (6)] Equation (6) contains a typographical error: the MSE term is printed with a misplaced closing parenthesis ('||...||)2 2'), and the notation 'R𝑡𝑖' for the true CTR vector is undefined. Please fix the formula and define all symbols.
  4. [§4.1, test set filtering] The paper does not define the 'relative CTR difference' used to set the D threshold (1% for training, 5% for testing). Please specify whether D is (CTR_high - CTR_low)/CTR_high, /CTR_low, or some other normalization, as this directly affects how interpretable the reported Pair Accuracy values are.
  5. [Appendix A.5] The listed limitation about aggregated user preferences is appreciated, but it does not address the more immediate distribution-shift concern discussed above: the RM's reliability on freshly generated images. Please add a limitation statement that acknowledges the need to validate the reward signal on the deployment distribution.

Circularity Check

0 steps flagged · score 1.0 of 10

No meaningful circularity: the reward model is trained on real click labels and the headline results are measured by online CTR, an external ground truth.

full rationale

The paper's derivation chain is externally anchored at both ends. The reward model (RM) is trained on historical advertising-image pairs with real user click data, using a binary cross-entropy loss plus a point-wise CTR regression loss (Eqs. 4-7), and its reported accuracy is measured on held-out test pairs filtered to require high exposure and a 5% CTR difference (Section 4.1). The DPO/PCPO stage (Eqs. 10-12) uses the RM's pairwise judgments as preference labels, but this is the standard RLHF proxy loop and not circular because the final evaluation is a real one-week online experiment with over 10 million impressions and a separate A/B test over 60 million impressions (Section 4.4). The match-rate metric is also assessed by human advertising professionals (Section 4.3.1), independent of the RM. There is no equation in which a predicted quantity is defined by construction in terms of the quantity it is supposed to predict. The paper contains several self-citations (e.g., [9], [23], [24], [44]) by overlapping authors, but these are used as background references for RLHF and existing generation pipelines, not as load-bearing justifications for the central claim. The strongest legitimate concern is a correctness/robustness issue rather than circularity: the RM is trained and tested on historical image pairs with large, clean CTR gaps, while Algorithm 1 applies it to freshly generated out-of-distribution image pairs whose true CTR differences may be small; if RM accuracy there is near chance, the DPO term would add noise. That is an empirical validity risk, not a self-referential reduction, and it does not make the online evaluation circular.

Assumptions & free parameters 12 free parameters · 4 assumptions · 0 invented entities

The system rests on standard machine learning assumptions (e.g., logged click data reflect user preference) plus several modeling choices specific to this paper: pairwise CTR as reward, fixed diffusion model, and threshold-based data filtering. The free parameters are mostly hyperparameters and data filters.

free parameters (12)
  • lambda_1 = 1
    Weight on the binary cross-entropy term in the RM loss (Eq. 7); chosen by hand.
  • lambda_2 = 0.5
    Weight on the point-wise CTR regression term in the RM loss (Eq. 7); chosen by hand.
  • E_train = 50
    Minimum exposure threshold for RM training pairs; filters noisy CTR estimates.
  • D_train = 1%
    Minimum relative CTR difference for RM training pairs.
  • E_test = 1000
    Exposure threshold for test pairs; makes test set easier, a selection-bias concern.
  • D_test = 5%
    CTR difference threshold for test pairs; test pairs are thus not representative of typical deployment pairs.
  • mask_ratio_visual = 75%
    PCPO visual-aware negative augmentation masks 75% of the product image (Section 3.4); chosen by hand.
  • beta_dpo = not specified
    DPO regularization coefficient in Eq. (10); the paper does not report the value.
  • learning_rate_pretrain = 2e-6
    Learning rate for e-commerce pre-training, reported in implementation details.
  • learning_rate_lora = 2e-5
    LoRA learning rate for preference optimization, reported in implementation details.
  • num_epochs_pretrain = 10
    Pre-training epochs.
  • num_epochs_rl = 5
    Preference optimization epochs.
assumptions (4)
  • domain assumption Pairwise CTR comparison is a valid reward signal for absolute CTR optimization
    Algorithm 1 uses RM pairwise rankings as the only reward; if pairwise order does not imply higher absolute CTR, the RL objective is misspecified. The online A/B test is the only check, and it is reported without significance.
  • domain assumption Fine-tuning only the Prompt Model is sufficient for CTR optimization
    Section 3.4: 'we empirically find that fine-tuning the background generation model has a much smaller impact... Therefore, we focus solely on fine-tuning the PM.' This claim is borrowed from Parrot and not re-validated in this paper.
  • domain assumption RM generalizes from filtered historical pairs to generated image pairs
    RM is trained on historical ads with exposure thresholds; at inference it scores freshly generated backgrounds, a distribution shift that the paper does not analyze.
  • domain assumption Inpainting with the product mask preserves product identity
    Eq. (9) assumes the masked diffusion update composites the product cleanly; if the mask leaks or causes artifacts, CTR comparisons would be measuring artifacts rather than background quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CTR-Driven Advertising Image Generation with Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/FM6PVKGA

@misc{pith2026250206823,
  author       = {Pith},
  title        = {Pith review of: CTR-Driven Advertising Image Generation with Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FM6PVKGA}},
  note         = {Machine review of arXiv:2502.06823}
}
read the original abstract

In web data, advertising images are crucial for capturing user attention and improving advertising effectiveness. Most existing methods generate background for products primarily focus on the aesthetic quality, which may fail to achieve satisfactory online performance. To address this limitation, we explore the use of Multimodal Large Language Models (MLLMs) for generating advertising images by optimizing for Click-Through Rate (CTR) as the primary objective. Firstly, we build targeted pre-training tasks, and leverage a large-scale e-commerce multimodal dataset to equip MLLMs with initial capabilities for advertising image generation tasks. To further improve the CTR of generated images, we propose a novel reward model to fine-tune pre-trained MLLMs through Reinforcement Learning (RL), which can jointly utilize multimodal features and accurately reflect user click preferences. Meanwhile, a product-centric preference optimization strategy is developed to ensure that the generated background content aligns with the product characteristics after fine-tuning, enhancing the overall relevance and effectiveness of the advertising images. Extensive experiments have demonstrated that our method achieves state-of-the-art performance in both online and offline metrics. Our code and pre-trained models are publicly available at: https://github.com/Chenguoz/CAIG.

Figures

Figures reproduced from arXiv: 2502.06823 by the authors.

Figure 1
Figure 1. (a) Example of the impact of different backgrounds on product CTR. While visual features play a crucial role, other [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. We first pre-train the MLLM on a large-scale multimodal [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 2
Figure 2. (a) E-commerce knowledge pre-training. The MLLM is pre-trained on a large-scale multimodal e-commerce dataset [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Comparison of Pair Accuracy across different methods on commercial and public datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 4
Figure 4. Figure 4: Comparison of Match Rate across different prefer [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison between DPO and the proposed PCPO. The first line shows the name of the product, followed by the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Advertising images generated by directly using the e-commerce knowledge-injected MLLM as PM. For each product, [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Some match and mismatch examples identified by annotators. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Extensive visual examples of our CAIG method applied to diverse product categories. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 24 canonical work pages

  1. [1]

    GPT-4V(ision) System Card

    2023. GPT-4V(ision) System Card. https://openai.com/index/gpt-4v-system- card/

  2. [2]

    Claude3.5-sonnet

    2024. Claude3.5-sonnet. https://www.anthropic.com/news/claude-3-5-sonnet

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

  4. [4]

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 (2022)

  5. [5]

    Binghui Chen, Chongyang Zhong, Wangmeng Xiang, Yifeng Geng, and Xuan- song Xie. 2024. VirtualModel: Generating Object-ID-retentive Human-object Interaction Image by Diffusion Model for E-commerce Marketing. arXiv preprint arXiv:2405.09985 (2024)

  6. [6]

    J Chen, J Xu, G Jiang, T Ge, Z Zhang, D Lian, and K Zheng. [n. d.]. Automated Creative Optimization for E-Commerce Advertising. arXiv 2021. arXiv preprint arXiv:2103.00436 ([n. d.])

  7. [7]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality. https://lmsys.org/blog/2023-03-30-vicuna/

  8. [8]

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. Advances in neural information processing systems 30 (2017)

Show all 59 references
  1. [9]

    Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al . 2025. Towards Reliable Ad- vertising Image Generation Using Human Feedback. In European Conference on Computer Vision. Springer, 399–415

  2. [10]

    Tiezheng Ge, Liqin Zhao, Guorui Zhou, Keyu Chen, Shuying Liu, Huimin Yi, Zelin Hu, Bochao Liu, Peng Sun, Haoyu Liu, et al. 2018. Image matters: Visually modeling user behaviors using advanced model server. In Proceedings of the 27th ACM International Conference on Information ...

  3. [11]

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, et al. 2024. ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools. arXiv preprint arXiv:2406.12793 (2024)

  4. [12]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial WWW ’25, April 28-May 2, 2025, Sydney, NSW, Australia Xingye Chen, et al. networks. Commun. ACM 63, 11 (2020), 139–144

  5. [13]

    Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. 2024. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024)

  6. [14]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

  7. [15]

    Jitesh Jain, Jianwei Yang, and Humphrey Shi. 2024. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 27992–28002

  8. [16]

    Chen Jie-Hao, Li Xue-Yi, Zhao Zi-Qian, Shi Ji-Yun, and Zhang Qiu-Hong. 2017. A CTR prediction method based on feature engineering and online learning. In2017 17th International Symposium on Communications and Information Technologies (ISCIT). IEEE, 1–6

  9. [17]

    Yuchin Juan, Yong Zhuang, Wei-Sheng Chin, and Chih-Jen Lin. 2016. Field- aware factorization machines for CTR prediction. In Proceedings of the 10th ACM conference on recommender systems . 43–50

  10. [18]

    Yueh-Ning Ku, Mikhail Kuznetsov, Shaunak Mishra, and Paloma de Juan. 2023. Staging e-commerce products for online advertising using retrieval assisted image generation. arXiv preprint arXiv:2307.15326 (2023)

  11. [19]

    Rohit Kumar, Sneha Manjunath Naik, Vani D Naik, Smita Shiralli, VG Sunil, and Moula Husain. 2015. Predicting clicks: CTR estimation of advertisements using logistic regression classifier. In 2015 IEEE international advance computing conference (IACC). IEEE, 1134–1138

  12. [20]

    Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. 2023. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)

  13. [21]

    Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Junfeng He, et al . 2024. Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation. arXiv preprint arXiv:2401.05675 (2024)

  14. [22]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2024. LLaVA-OneVision: Easy Visual Task Transfer. arXiv preprint arXiv:2408.03326 (2024)

  15. [23]

    Fengheng Li, An Liu, Wei Feng, Honghe Zhu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, Zhangang Lin, et al. 2023. Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the 32nd ACM International Conference on Information and ...

  16. [24]

    Zhaochen Li, Fengheng Li, Wei Feng, Honghe Zhu, An Liu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, et al. 2023. Planning and Rendering: Towards End-to-End Product Poster Generation. arXiv preprint arXiv:2312.08822 (2023)

  17. [25]

    Kaiyi Lin, Xiang Zhang, Feng Li, Pengjie Wang, Qingqing Long, Hongbo Deng, Jian Xu, and Bo Zheng. 2022. Joint Optimization of Ad Ranking and Creative Se- lection. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval...

  18. [26]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proce...

  19. [27]

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024. LLaVA-NeXT: Improved reasoning, OCR, and world knowl- edge. https://llava-vl.github.io/blog/2024-01-30-llava-next/

  20. [28]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024. Visual instruc- tion tuning. Advances in neural information processing systems 36 (2024)

  21. [29]

    Yanqing Liu, Kai Wang, Wenqi Shao, Ping Luo, Yu Qiao, Mike Zheng Shou, Kaipeng Zhang, and Yang You. 2023. Mllms-augmented visual-language repre- sentation learning. arXiv preprint arXiv:2311.18765 (2023)

  22. [30]

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. 2022. Repaint: Inpainting using denoising diffusion proba- bilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11461–11471

  23. [31]

    James MacGlashan, Mark K Ho, Robert Loftin, Bei Peng, Guan Wang, David L Roberts, Matthew E Taylor, and Michael L Littman. 2017. Interactive learning from policy-dependent human feedback. In International conference on machine learning. PMLR, 2285–2294

  24. [32]

    Shaunak Mishra, Manisha Verma, Yichao Zhou, Kapil Thadani, and Wei Wang

  25. [33]

    Phillip Mueller, Jannik Wiese, Ioan Craciun, and Lars Mikelsons. 2024. InsertD- iffusion: Identity Preserving Visualization of Objects through a Training-Free Diffusion Architecture. arXiv preprint arXiv:2407.10592 (2024)

  26. [34]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35...

  27. [35]

    Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. 2024. Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning. arXiv preprint arXiv:2408.10075 (2024)

  28. [36]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

  29. [37]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems 36 (2024)

  30. [38]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

  31. [39]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  32. [40]

    Anand Siththaranjan, Cassidy Laidlaw, and Dylan Hadfield-Menell. 2023. Distri- butional preference learning: Understanding and accounting for hidden context in RLHF. arXiv preprint arXiv:2312.08358 (2023)

  33. [41]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

  34. [42]

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. 2020. Learning to summarize with human feedback. Advances in Neural Information Processing Systems 33 (2020), 3008–3021

  35. [43]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

  36. [44]

    Haohan Wang, Wei Feng, Yang Lu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, Zhangang Lin, Lixing Bo, et al. 2025. Generate E-commerce Product Background by Integrating Category Commonality and Personalized Style. In ICASSP 2025-2025 IEEE International Conference ...

  37. [45]

    Shiyao Wang, Qi Liu, Tiezheng Ge, Defu Lian, and Zhiqiang Zhang. 2021. A hybrid bandit model with visual priors for creative ranking in display advertising. In Proceedings of the web conference 2021 . 2324–2334

  38. [46]

    Shiyao Wang, Qi Liu, Yicheng Zhong, Zhilong Zhou, Tiezheng Ge, Defu Lian, and Yuning Jiang. 2022. CreaGAN: An Automatic Creative Generation Framework for Display Advertising. In Proceedings of the 30th ACM International Conference on Multimedia. 7261–7269

  39. [47]

    Penghui Wei, Shaoguo Liu, Xuanhua Yang, Liang Wang, and Bo Zheng

  40. [48]

    Penghui Wei, Xuanhua Yang, Shaoguo Liu, Liang Wang, and Bo Zheng. 2022. CREATER: CTR-driven advertising text generation with controlled pre-training and contrastive fine-tuning. arXiv preprint arXiv:2205.08943 (2022)

  41. [49]

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. 2023. Bet- ter aligning text-to-image models with human preference. arXiv preprint arXiv:2303.14420 1, 3 (2023)

  42. [50]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 (2024)

  43. [51]

    Hao Yang, Jianxin Yuan, Shuai Yang, Linhe Xu, Shuo Yuan, and Yifan Zeng. 2024. A New Creative Generation Pipeline for Click-Through Rate with Stable Diffusion Model. In Companion Proceedings of the ACM on Web Conference 2024 . 180–189

  44. [52]

    Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. 2024. Contextual object detection with multimodal large language models.International Journal of Computer Vision (2024), 1–19

  45. [53]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional con- trol to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision . 3836–3847

  46. [54]

    Kang Zhao, Xinyu Zhao, Zhipeng Jin, Yi Yang, Wen Tao, Cong Han, Shuanglong Li, and Lin Liu. 2024. Enhancing Baidu Multimodal Advertisement with Chinese Text-to-Image Generation via Bilingual Alignment and Caption Synthesis. In Proceedings of the 47th International ACM SIGIR Co...

  47. [55]

    Zhanhui Zhou, Jie Liu, Chao Yang, Jing Shao, Yu Liu, Xiangyu Yue, Wanli Ouyang, and Yu Qiao. 2023. Beyond one-preference-for-all: Multi-objective direct prefer- ence optimization. arXiv preprint arXiv:2310.03708 (2023)

  48. [56]

    Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593 (2019). CTR-Driven Advertising Image Generation with Multimoda...

  49. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

  50. [2020]

    In Proceedings of the 29th ACM international conference on information & knowledge management

    Learning to create better ads: Generation and ranking approaches for ad creative refinement. In Proceedings of the 29th ACM international conference on information & knowledge management . 2653–2660

  51. [2022]

    In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Towards personalized bundle creative generation with contrastive non- autoregressive decoding. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 2634–2638

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.