acceptodds
Under review as a conference paper at ICLR 2027

DiffAuction: An Auction-Driven Framework for Generative Image Advertising

Abstract

Diffusion models create an opportunity for advertising in generated visual content, with multiple advertisers competing for exposure within a shared image. Yet existing approaches are computationally costly when generating images under varying target-share configurations, and it remains unclear how the resulting outcomes should be evaluated and priced. To address these gaps, we propose DiffAuction, an auction framework connecting controllable generation, click-intent evaluation, and auction design. First, our training-free Residual Share Calibration (RSC) decomposes joint-attention updates into product-specific residuals and rescales them according to target shares. Because a fixed prompt tends to produce similar viewpoints that favor the same products regardless of their target shares, we further introduce Share-Conditioned Prompting (SCP) to adapt the viewpoint to each desired allocation. Second, we develop an evaluator that uses a VLM to extract visual evidence and an LLM to assess Click-Through Intent (CTI), a proxy for click propensity, along four dimensions: Product Prominence, Brand Recognizability, Scene Integration, and Contextual Value Communication. We validate its assessments against judgments from 10 trained annotators on 1,000 images. Third, we formulate this advertising setting as an auction over jointly feasible generation configurations. Each configuration has a public expected CTI vector, while advertisers hold configuration-specific private values per unit of CTI that map click intent into commercial value. We adopt a learnable affine maximizer auction (AMA), which jointly selects a shared generation configuration and determines advertiser-specific payments to maximize expected platform revenue. The resulting allocation and payment rules guarantee dominant-strategy incentive compatibility (DSIC) and individual rationality (IR). Experiments show that our generation method strengthens the monotonic relationship between target shares and CTI relative to baselines, while the learned AMA achieves higher platform revenue in the resulting auction setting.

Primary area:
generative models
Keywords:
Diffusion Models, Generative Advertising, Auction Mechanism Design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZxxsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Overall, this is an industry application focussed paper where the authors propose a framework for ad image generation, performance estimation, and auction to optimize for ad revenue. The author's clearly outline the challenges in this setup (i.e., controlling the share of each brand in terms of real estate/prominence in the image, estimating clicks, and optimizing revenue). The proposed approach includes heuristics in the prompt for image generation (making it senstive to target shares), a VLM evaluator for click propensity, and an auction mechanism in sync with their setup (showing theoretical guarantees, e.g., dominant strategy incentive compatibility and individual rationality). The authors share postiive results on synthetic data, validating claims around revenue maximization. In terms of shortcomings, the paper is limited in terms of (1) modeling contributions (mostly has heuristics and straightforward LLM/VLM applications except the auction design part where proving guarantees is not trivial) and (2) validation on real advertising data/acceptance from advertisers and ad platforms.

Critical strengths
  1. One of the core strengths in this paper is the clean decomposition of challenges in a real world product setting. The proposed framework has separate blocks for share sensitive prompting, VLM based click propensity estimation and auction design for revenue maximization.

  2. Authors have shared theoretical guarantees around dominant strategy incentive compatibility and individual rationality

  3. Positive results on synthetic data.

Critical weaknesses
  1. Modeling novelty/contributions are limited. Except the auction design part, the framework mostly involves prompting heuristics and direct applications of existing models.

  2. Experiments are on synthetic data. There is no evidence around how well it will work on real world advertising data, and if advertisers/ad platforms will actually adopt this approach.

  3. Real world ad platforms and advertiser expectations are more complex than in the paper's setting. For example, there may be errors in click propensity estimation or an advertiser may not get the desired ROI. How will the proposed framework handle such situations - this is where experiments in an actual ad platform would have meaningfully strengthened the paper.

Questions for authors
  1. Have the authors experimented with any datasets from real advertising platforms?
4 comments

Thank you for recognizing the clear decomposition of the problem and the incentive guarantees provided by our work. Below, we will address your questions point by point.

For Weakness 1:

Modeling novelty/contributions are limited. Except the auction design part, the framework mostly involves prompting heuristics and direct applications of existing models.

We would like to clarify that our generation method involves more than prompt engineering: it also intervenes in the internal computational process of diffusion models. Residual Share Calibration (RSC) decomposes the joint-attention outputs into product-specific residual contributions and normalizes and reweights their magnitudes according to the target exposure shares. This process requires only a single model forward pass per denoising step. Share-Conditioned Prompting (SCP) complements this internal control mechanism by adjusting viewpoints and composition.

Beyond intervening in the generation process, the combination of RSC and SCP provides effective complementary control. RSC is effective at responding to continuous changes in exposure shares and adjusting product exposure accordingly. However, the generated viewpoints remain constrained by the prompts. We therefore introduce SCP, which assigns different prompts to different share configurations, overcoming the limitations imposed by fixed viewpoints on share control.

We conduct controlled ablation studies to distinguish the contributions of RSC from those of prompt-based strategies. With the same SCP configuration, RSC + SCP achieves a Spearman correlation of 0.8346 between target exposure shares and CTI, outperforming the baseline method + SCP (0.7837, Table 2). RSC also outperforms direct residual scaling, achieving a correlation of 0.8346 compared with 0.7705 (Table 12). These results demonstrate that the performance improvements cannot be attributed solely to prompt engineering.

Furthermore, across the evaluated numbers of products, our method achieves 1.91–4.58× faster per-step inference than the baseline (Table 3).

Our contributions also include turning generated images into allocatable auction items and designing corresponding CTI estimation and pricing mechanisms. While prior work has primarily studied auctions in LLM settings, we focus on auctions in image generation and address controllable generation, generation efficiency, CTI estimation, and pricing. We believe these contributions help advance research in this field.

For Weakness 2 & 3:

Experiments are on synthetic data. There is no evidence around how well it will work on real world advertising data, and if advertisers/ad platforms will actually adopt this approach.

Thank you for your suggestion. We acknowledge the importance of real-world platform validation and would like to clarify the research positioning and empirical foundations of our work.

First, validating mechanisms through constructed scenarios, simulated feedback, or model-based evaluation is an approach widely adopted by related studies. For example, Dütting et al. (2024) and MOSAIC (Soumalias et al., 2025) use fictitious advertisers or synthetic instances. Sun et al. (2025) evaluate image auctions using constructed prompt scenarios and CLIP alignment scores. LLM-Auction (Zhao et al., 2025), although using real advertising materials, still relies on simulated click feedback and bids. IAMFM (Liu et al., 2026) and Neuron Auctions (Yun et al., 2026) also employ model-based evaluations of utility or click intention. Turner-Smith et al. (2026) explicitly identify the unavailability of behavioral logs and use agent-based simulations to construct click-intention supervision. Since these business models are still emerging, existing platform data often cannot provide actual feedback, such as CTR and conversion rates, for specific combinations of generated content and allocation mechanisms.

Therefore, few studies in this field use real-world platform data, but this does not diminish their value. Rather, these studies offer diverse approaches and insights into emerging economic models in the AI era, making them valuable research contributions.

Consequently, these studies primarily investigate how generative capabilities can be translated into allocatable and priceable auction items, establishing methodological foundations for practical deployment.

Our work focuses on transforming product exposure in generated images into controllable, evaluable, and priceable auction items, which is a fundamental prerequisite for implementing this business model. We achieve controllable product exposure generation through the combination of RSC and SCP, and verify the monotonic relationship between target exposure shares and CTI. In the absence of real-world advertising CTR data, we further compare predicted CTI with independently reported human click intentions and observe an obvious correlation, providing human-based evidence for evaluating the effectiveness of generated images. Building on this, we develop a configuration-based auction mechanism with DSIC and IR guarantees, providing a concrete implementation for the allocation and pricing of shared generated images. Together, these experiments and theoretical analyses provide preliminary evidence for the feasibility of introducing image generation into auction-based advertising. The proposed exposure control method, CTI evaluation metric, and auction mechanism provide methodological guidance and technical foundations for future practical applications.

Online validation of this emerging advertising format requires not only building a platform that supports shared image generation and auctions, but also recruiting real advertisers and conducting actual advertising campaigns. Such validation cannot be readily achieved through existing platform APIs or historical logs. It requires substantial infrastructure development and commercial partnerships, which are beyond the scope of what we can currently undertake. Therefore, we first focus on establishing and validating the key methods necessary to support this business model.

For Question 1:

Have the authors experimented with any datasets from real advertising platforms?

To evaluate our method on real user requests, we selected 100 prompts suitable for product placement from DiffusionDB. We preserved the original scene descriptions and paired each scene with two contextually relevant branded products. We then generated images using RSC and Share-Conditioned Prompting (SCP) under different target shares and evaluated the monotonic relationship between target shares and CTI.

Evaluation SettingPearson ↑Spearman ↑Pairwise ↑Kendall’s τb\tau_b ↑Isotonic R2R^2 ↑
Real user prompts from DiffusionDB (100 prompts)0.74060.77760.80290.63000.8485

These results indicate that our method maintains effective share control in test scenarios derived from real user prompts.

Furthermore, when predicting CTI, we asked multiple annotators to provide scores based on their actual willingness to click and evaluated the agreement between model predictions and human ratings, with the results reported in Appendix 3. We therefore validated, to the extent feasible, the alignment between model predictions and human preferences, and evaluated share control using real user prompts.

We hope these explanations can address your concerns and sincerely look forward to your response.

Dear Reviewer Zxxs,

Thank you again for your thoughtful feedback. We would appreciate the opportunity to hear your thoughts on our responses and additional evaluation. To briefly summarize:

  • W1: We clarified how RSC intervenes in the diffusion model’s internal computation and highlighted controlled ablations demonstrating its contribution beyond prompting, alongside the reported 1.91–4.58× per-step inference speedup.
  • W2 and W3: We explained the methodological scope of our work, discussed the challenges of validating this emerging advertising format on deployed platforms, and clarified the evidence provided by our controlled experiments and human click-intent evaluation.
  • Q1: We added an evaluation using 100 real user prompts from DiffusionDB, showing that share control remains effective in these scenarios. It extends our evaluation to scenes derived from real user requests.

We would appreciate your feedback on whether our clarifications and additional evaluation address your main concerns. If any concerns remain, it would be helpful to know which aspects require further discussion so that we can respond more precisely. We hope you will consider this additional evidence when reassessing our work.

Thank you for your time and consideration.

Best regards,
The Authors

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

The paper proposes a framework for allocating advertising exposure among products in a shared generated image. Residual Share Calibration (RSC) rescales product-associated joint-attention residuals according to target shares, while Share-Conditioned Prompting (SCP) adjusts prominence instructions. A VLM-LLM pipeline estimates product-level click-through intent (CTI) using four visual criteria. Historical CTI averages then define a public configuration table for a learned affine maximizer auction, with configuration-specific private valuations and payments satisfying DSIC and IR under the stated valuation model. Experiments evaluate share-CTI monotonicity, generation speed, agreement with human judgments, and simulated auction revenue.

Critical strengths
  • S1. Practical improvement in controllable generation. RSC offers a simple, training-free intervention with one model forward pass per denoising step. The comparison with an SCP-enhanced baseline is particularly useful. On the 200-scene ablation set, RSC + SCP improves Spearman correlation from 0.7837 to 0.8346 and pairwise monotonicity from 0.7797 to 0.8024 (Table 2). Table 3 also reports per-timestep speedups of 1.91× to 4.58× over the paper's implementation of the Sun et al. (2025) baseline as the number of products increases from two to ten. These results support a useful efficiency and control contribution.

  • S2. Meaningful human validation of the evaluator. The paper provides both rubric-based annotations on 1,000 images and a separate study of self-reported willingness to click. In Appendix C, five additional annotators evaluate 570 images from 30 scenes without seeing the rubric, yielding Pearson/Spearman correlations of 0.8526/0.8339 with predicted CTI. This additional study strengthens the case that the evaluator captures an aspect of human interest beyond reproducing its own scoring instructions.

Critical weaknesses

W1. The revenue experiments do not isolate the economic benefit of the proposed integration

  • Evidence: Section 3.3 adopts an existing affine maximizer construction. Learning AMA parameters with a differentiable selection surrogate is already established in Section 5 of Curry et al. (2023), which DiffAuction itself cites. In Section 4.3 and Table 6, the CTI table is fixed, and the learned mechanism is compared with VCG and a Myerson-based mechanism with fixed weights and offsets. I could not find a comparison using the same learned auction with configuration tables produced by alternative generators, or an ablation separating learned offsets from learned advertiser weights. The approximately 51% revenue gains over the Myerson-based baseline occur when the configuration-specific values per CTI unit are sampled independently. Relative to that same baseline, the paper reports gains of 5.10% and 4.27% under smoothly correlated values. Appendix D reports 131,072 independent test valuation profiles per setting, but this tests generalization within the stipulated valuation distributions.
  • Impact on recommendation: The results establish that fitting AMA parameters improves revenue over the selected fixed mechanisms. They leave unresolved how much economic benefit follows from the new generation and evaluation components. Since integration with auction design is a central contribution, a controlled comparison connecting those components to the economic outcome would materially strengthen the significance of the work. A simpler learned comparator would also clarify which mechanism parameters account for the gains.

W2. The validation leaves a gap between image-level CTI agreement and the public table used for allocation

  • Evidence: Section 1 frames the auction for a given scene and set of competing advertisers (lines 77-84). Equation (6), however, estimates CTI by averaging over historical scenes, and Section 4.3 uses this fixed table across valuation profiles. Appendix G constructs scenes with changing product identities, leaving the correspondence between those scene-product records and the advertiser index in the auction table unclear. I could not find a specification of whether the experimental table is estimated for fixed advertiser/product identities, generic product positions, or separately conditioned scene groups, or a test of its CTI prediction on held-out scenes for the relevant advertiser set. Appendix C is encouraging, but its reported correlations summarize 570 images from 30 scenes, with 19 generated images per scene. It does not report human-model ranking agreement within each scene-product curve or uncertainty accounting for the shared scenes. Thus, the reported aggregate agreement does not establish the precision needed to distinguish configurations for a particular advertising instance.
  • Impact on recommendation: The CTI table connects generation to allocation and pricing, so its construction and predictive reliability are central to interpreting the system’s results. Appendix E.4 appropriately states that historical CTI alone does not establish equality with true expected advertiser value. I do not see this as a flaw in the conditional DSIC/IR proof. However, the current evidence leaves the practical usefulness of the auction’s public inputs insufficiently established. Clarifying the table construction and validating configuration rankings on held-out scenes would address this concern without requiring a large online advertising trial.
Questions for authors
  1. To clarify W1, how much of the revenue improvement remains with unit advertiser weights and learned configuration offsets only? Under a common valuation protocol, how does the same learned AMA perform with configuration tables from RSC + SCP versus the SCP-enhanced Sun et al. baseline? These comparisons would clarify the contribution of the auction parameters and whether the generation improvement translates into economic benefit.

  2. To clarify W2, what exactly does advertiser index ii represent when constructing qikq_{ik} across historical scenes with different products? Please specify the conditioning and estimation procedure, including the 19 three-advertiser configurations. Can you report held-out prediction error for the resulting CTI table and within-scene-product human-model rank agreement from Appendix C, with uncertainty estimated across scenes? Evidence of reliable configuration discrimination would substantially reduce this concern.

if author can adress those two question. I will be happy to raise my score. The performance of the method is proven to be powerful.

6 comments

Thank you for recognizing the merits of our work and for your valuable and constructive feedback. We appreciate the opportunity to clarify our work and address your concerns in detail below.

For Question 1:

How much of the revenue improvement remains with unit advertiser weights and learned configuration offsets only?

We find that the contribution of advertiser weights depends on whether advertisers have identical distributions of private values.

The two valuation distributions used in our paper do not account for distributional differences across advertisers. Under these settings, fixing all advertiser weights to 1 has almost no effect on the final revenue.

We further investigate several settings in which advertisers have different valuation distributions and calculate the gain retention rate.

We define the gain retention rate as

GainRetention=Rbias−only−RfixedRfull−Rfixed×100%.\mathrm{GainRetention}=\frac{R_{\mathrm{bias-only}}-R_{\mathrm{fixed}}}{R_{\mathrm{full}}-R_{\mathrm{fixed}}}\times 100\%.

Here, RfixedR_{\mathrm{fixed}} denotes the revenue of the reference AMA with advertiser weights fixed at 1 and configuration biases fixed as follows:

$ w_i=1,\qquad \beta_k=-\frac{1}{2}\sum_{i=1}^{n}q_{ik},\qquad \beta_0=0.$

Under these settings, advertiser weights play a more significant role, as shown in the table below. The entries report gain retention rates with 95% confidence intervals.

Valuation SettingTwo AdvertisersThree Advertisers
Independent sampling / Different value ranges96.49% [95.95, 97.03]93.16% [92.16, 94.17]
Independent sampling / High- and low-value types85.43% [85.10, 85.76]85.14% [84.63, 85.64]
Independent sampling / Equal means, different shapes88.86% [88.38, 89.34]86.74% [86.04, 87.43]
Correlated configurations / High- and low-value types79.83% [79.05, 80.61]63.49% [61.61, 65.37]
Endpoint interpolation / High- and low-value types98.27% [97.64, 98.89]89.87% [88.30, 91.45]

These analyses show that when valuation distributions differ across advertisers, learning advertiser weights provides additional benefits, although configuration offsets account for most of the revenue gains. When advertisers share the same valuation distribution, fixing all weights at 1 yields nearly the same performance. Thus, learnable advertiser weights help accommodate a broader range of valuation settings, while learning only the offsets retains nearly all the performance when advertisers’ valuation distributions are similar.

The different settings are described below:

Valuation SettingDistribution DesignInterpretation
Independent sampling / Different value rangesTwo advertisers: U(0,1)U(0,1) and U(0,1/4)U(0,1/4). Three advertisers: U(0,1)U(0,1), U(0,1/2)U(0,1/2), and U(0,1/4)U(0,1/4). Values are sampled independently across configurations.Advertisers differ in their willingness-to-pay levels.
Independent sampling / High- and low-value typesTwo advertisers: Beta(1,1/2)\mathrm{Beta}(1,1/2) and Beta(1,4)\mathrm{Beta}(1,4). Three advertisers: Beta(1,1/2)\mathrm{Beta}(1,1/2), U(0,1)U(0,1), and Beta(1,4)\mathrm{Beta}(1,4). Values are sampled independently across configurations.Some advertisers typically have high unit values, while others typically have low unit values.
Independent sampling / Equal means, different shapesTwo advertisers: Beta(2,2)\mathrm{Beta}(2,2) and Beta(1/2,1/2)\mathrm{Beta}(1/2,1/2). Three advertisers: Beta(2,2)\mathrm{Beta}(2,2), U(0,1)U(0,1), and Beta(1/2,1/2)\mathrm{Beta}(1/2,1/2). Values are sampled independently across configurations. All distributions have mean 0.5 and support [0,1].Advertisers have the same average willingness to pay, but their values concentrate around the middle or near the two extremes.
Correlated configurations / High- and low-value typesTwo advertisers have marginals Beta(1,1/2)\mathrm{Beta}(1,1/2) and Beta(1,4)\mathrm{Beta}(1,4); three advertisers have marginals Beta(1,1/2)\mathrm{Beta}(1,1/2), U(0,1)U(0,1), and Beta(1,4)\mathrm{Beta}(1,4). Within each advertiser, configuration values follow a Gaussian copula with latent correlation ρ=0.75\rho=0.75.An advertiser’s unit values across configurations tend to be high or low together.
Endpoint interpolation / High- and low-value typesEach advertiser independently draws two endpoint unit values from its corresponding distribution in the second row, then linearly interpolates between them over its own candidate target-share range. The endpoints are not sorted.An advertiser’s unit value changes linearly with its own target share and may increase or decrease.

Under a common valuation protocol, how does the same learned AMA perform with configuration tables from RSC + SCP versus the SCP-enhanced Sun et al. baseline?

Following the reviewer’s suggestion, we compare the economic outcomes achieved by the enhanced baseline and our method. To assess the robustness of the revenue gains across valuation settings, we broaden our evaluation to include common valuation distributions informed by those used in the economics literature. For each setting, we report the percentage revenue improvement of our method over the enhanced baseline:

Valuation SettingTwo AdvertisersThree Advertisers
Endpoint interpolation+1.18% [1.06, 1.31]−0.41% [−0.59, −0.24]
Independent uniform+0.48% [0.35, 0.62]+2.34% [2.12, 2.56]
Correlated uniform+2.31% [2.17, 2.45]+1.98% [1.80, 2.16]
Beta(1,2)+2.78% [2.63, 2.92]+4.18% [3.98, 4.38]
Beta(2,1)+2.37% [2.24, 2.50]+1.60% [1.43, 1.77]
Beta(2,2)+2.09% [1.96, 2.21]+2.07% [1.90, 2.25]
Beta(1/2,1/2)+2.14% [2.01, 2.26]+1.91% [1.72, 2.10]
Truncated exponential+2.91% [2.74, 3.08]+4.98% [4.77, 5.20]
Truncated lognormal+2.99% [2.84, 3.13]+4.63% [4.41, 4.85]
Random ideal share: Ti∼U(0,1)T_i\sim U(0,1)+3.37% [3.15, 3.59]+4.73% [4.50, 4.95]
Independent sampling / Different value ranges+1.97% [1.89, 2.06]+1.93% [1.77, 2.09]
Independent sampling / High- and low-value types+1.84% [1.75, 1.93]+0.62% [0.47, 0.76]
Independent sampling / Equal means, different shapes+1.46% [1.34, 1.58]+3.43% [3.21, 3.64]
Correlated configurations / High- and low-value types+1.89% [1.79, 1.98]+3.32% [3.14, 3.49]
Endpoint interpolation / High- and low-value types+1.05% [0.89, 1.20]−0.47% [−0.63, −0.31]

These valuation settings cover several representative distribution types and consider both identical and heterogeneous valuation distributions across advertisers. The results show that our method achieves higher revenue in the vast majority of the evaluated settings.

Furthermore, compared with baseline, our method offers improvements in both exposure-share control and inference speed: our method achieves 1.91–4.58× faster per-step inference than the baseline (Table 3).

The different settings are described below:

Valuation SettingDistribution DesignInterpretation
Endpoint interpolationEach advertiser independently draws two endpoints from U(0,1)U(0,1) and linearly interpolates between them over its own candidate target-share range. The endpoints are not sorted.Unit values change gradually with target share and may increase or decrease.
Independent uniformUnit values are sampled independently from U(0,1)U(0,1) across configurations.A basic multidimensional benchmark with independent values across configurations.
Correlated uniformEach configuration has a U(0,1)U(0,1) marginal. Within each advertiser, configuration values follow a Gaussian copula with latent correlation ρ=0.75\rho=0.75.An advertiser’s unit values across configurations tend to be high or low together.
Beta(1,2)Unit values have Beta(1,2)\mathrm{Beta}(1,2) marginals, with mean 1/31/3.Low willingness to pay is more common than high willingness to pay.
Beta(2,1)Unit values have Beta(2,1)\mathrm{Beta}(2,1) marginals, with mean 2/32/3.High willingness to pay is more common than low willingness to pay.
Beta(2,2)Unit values have Beta(2,2)\mathrm{Beta}(2,2) marginals, with mean 0.5 and greater concentration around the middle.Moderate willingness to pay is common, while extreme values are less frequent.
Beta(1/2,1/2)Unit values have Beta(1/2,1/2)\mathrm{Beta}(1/2,1/2) marginals, with mean 0.5 and greater concentration near both endpoints.Willingness to pay is polarized, with more values near zero or one.
Truncated exponentialX∼Exp(1)X\sim\mathrm{Exp}(1), conditioned on X∈[0,1]X\in[0,1].Low unit values are more common, with frequency decreasing as value increases. The standard rate 1 provides a simple low-value-concentrated setting.
Truncated lognormalln⁡X∼N(0,1)\ln X\sim\mathcal{N}(0,1), conditioned on X∈[0,1]X\in[0,1].A standard lognormal distribution restricted to [0,1], representing asymmetric unit values without introducing additional location or scale choices.
Random ideal shareTi∼U(0,1)T_i\sim U(0,1), with unit value Xikexp⁡[−4ln⁡2(sik−Ti)2]X_{ik}\exp[-4\ln 2(s_{ik}-T_i)^2].Each advertiser has an individual preferred share. Unit value decreases with distance from that share.

For Question 2:

what exactly does advertiser index i represent when constructing qik across historical scenes with different products? Can you report held-out prediction error for the resulting CTI table

qikq_{ik} denotes the expected CTI for product slot i when generation uses the target-share vector specified by configuration kk. We estimate it by averaging the predicted CTI scores of all products occupying slot ii across historical scenes under configuration kk. The resulting CTI table is not conditioned on product category or scene; instead, it maps target-share vectors to expected CTI vectors. For example, “Product A” denotes a product slot rather than a fixed product identity, and we average the CTI scores of all products occupying that slot across historical scenes.

We additionally conduct an experiment comparing two CTI estimation methods:

  1. Averaging by product slot and target share, as adopted in our paper, which aggregates CTI scores across training scenes for the same prompt position (A or B) and target share, regardless of product identity.
  2. Averaging by product identity and target share, which aggregates historical CTI scores only for the same product identity at the same target share. Product identity is determined by the specific brand and product description.

For this experiment, we construct 100 two-advertiser scenes, generating 19 images per scene, one for each target-share configuration. The scenes cover 20 distinct products, with each product appearing in 10 scenes. This allows us to investigate whether averaging historical CTI by product identity improves CTI prediction accuracy in new scenes.

We use five-fold cross-validation. In each fold, we use 80 scenes to construct the CTI tables for both estimation methods and evaluate prediction errors on the remaining 20 held-out scenes. We compute MAE between the CTI estimated from the historical lookup table and the CTI predicted directly from the generated images in the held-out scenes.

CTI table estimation methodMAE ↓ (95% CI)
Averaging by product slot0.1674 [0.1597, 0.1753]
Averaging by product identity0.1590 [0.1516, 0.1666]

We further evaluate the two CTI estimation methods on our original dataset of 1,000 scenes. According to our statistics, each product appears in 3–6 scenes. We conduct five-fold cross-validation with scene-level splits, obtaining MAEs of 0.1752 [0.1689, 0.1824] for slot-level aggregation and 0.1810 [0.1761, 0.1883] for product-specific aggregation, respectively. In this setting, slot-level aggregation achieves better performance.

We believe that when sufficient historical observations are available for each product, product-specific CTI estimation can better preserve product-specific characteristics, leading to more accurate estimates. However, when a product appears only a limited number of times in historical data, its estimated CTI may be biased toward the particular scenes in which it appeared, potentially reducing its generalization accuracy in new scenes.

Considering that products participating in real-world auctions may not have sufficient historical observations, product-specific estimation cannot always rely on adequate historical data. Therefore, we adopt the simpler slot-level aggregation strategy, which estimates the average CTI that the generation model produces for each product slot under different target-share configurations.

We acknowledge that the slot-level CTI estimation adopted in our paper is a simplified approximation. In practical applications, platforms could develop more fine-grained CTI estimation models by incorporating richer product-level historical data, scene characteristics, and other relevant factors to improve estimation accuracy and practical utility. Such improvements are compatible with our proposed framework and would not require modifications to its core allocation and payment rules.

Reviewer_ZZ1HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies exposure control, click-through intent estimation, and auction pricing when multiple advertised products share a generated image. Residual Share Calibration (RSC) decomposes product-specific residuals within a single joint-attention computation and calibrates their norms according to target shares. Share-Conditioned Prompting (SCP) uses five prominence levels to influence composition and viewing angles. The paper then constructs a product-level click-through intent (CTI) proxy through VLM-based visual evidence extraction and LLM-based scoring, and learns an affine maximizer auction over configurations and an expected-CTI table fixed in advance. Experiments report better share–CTI monotonicity and lower inference time per denoising timestep than the baseline, alongside human evaluations and revenue comparisons under synthetic private valuations. The DSIC and IR results apply under the stated valuation model and the hard-selection rule used at deployment.

Critical strengths
  1. The ablation and control experiments for exposure-share generation are thorough. Table 2 equips the baseline with the same share-conditioned prompting, showing that the gain is not attributable to prompt conditioning alone, and Table 12 compares RSC against directly scaling residuals by the target shares. Appendices A and B further examine the number of controlled denoising steps, the choice of FLUX transformer streams, settings with up to ten competing products, and competing products from the same category.

  2. The human validation of the CTI evaluator is well considered. Beyond the 1,000-image rubric-based evaluation in Section 4.2, Appendix C recruits five additional annotators without access to the four-dimensional rubric to rate click intent on 570 images, supporting CTI as a proxy for average self-reported click intent within the evaluated sample.

Critical weaknesses
  1. The human validation of CTI reports correlations but does not separately report ordering consistency as the target share varies for the same product within the same scene. That within-condition ordering is what the generation experiments actually evaluate, yet Appendix C gives a single set of correlations without stating whether they are computed per scene-product curve and then averaged, or pooled across all evaluation items. If these correlations are pooled, they also reflect variance across scenes and products, and so cannot on their own verify whether the share-induced ordering agrees with human click intent. Please specify the statistical unit and report human-model ranking agreement across share configurations within each scene-product pair.

  2. In the auction experiments, it is unclear how the public CTI table q_{ik} is mapped to advertiser identities. Equation (6) defines q_{ik} per advertiser i, whereas the dataset described in Appendix G spans 400 distinct products across 1,000 scenes. Section 4.3 says expected CTI is averaged over historical scenes, but does not explain whether q_{ik} is averaged by product identity, by product position within a scene (e.g., Product A vs. Product B), or over fixed product groups. While the DSIC proof holds for any fixed public table, leaving this mapping unspecified makes the auction experiments hard to reproduce or interpret. Please provide the public q table used in Table 6 and explain how it is built: what each advertiser index i corresponds to, and which scenes are averaged. Section 4.3 states that both settings use 19 paid configurations plus a no-allocation option, and Section 4.1 gives the two-advertiser grid from (0.95, 0.05) to (0.05, 0.95); please also state how the 19 configurations are constructed in the three-advertiser setting.

Questions for authors
  1. The tables report monotonicity only for the overall CTI score. How do the four individual dimensions behave as the target share increases, and if the per-dimension scores were retained, can you report their monotonicity?

  2. Appendix B.2 defines the five metrics on a single scene-product curve, and Section 4.1 states that the Figure 2 curves are aggregated over 1,000 scenarios. Please describe how the curve-level metrics are aggregated in Tables 1 and 2 and in the appendix tables, how constant or degenerate curves are handled for metrics such as Spearman correlation and Isotonic R^2, and the uncertainty around these aggregated values.

4 comments

Thank you for your careful review. We provide clarifications and additional results below to address your concerns and questions.

For Weakness 1:

Appendix C gives a single set of correlations without stating whether they are computed per scene-product curve and then averaged, or pooled across all evaluation items.

Thank you for highlighting the importance of within-condition ordering consistency. The correlations reported in Appendix C are computed within each scene–product pair, rather than pooled across different scenes or products.

Specifically, our evaluation contains 30 scenes, each with two products evaluated under 19 target-share configurations, yielding 60 scene–product curves. For each curve, we first average the click-intent ratings of five independent annotators at each configuration, and then compute the correlation between the resulting human-rating curve and the corresponding model-predicted CTI curve. The reported results are obtained by averaging these curve-level correlations across scene–product pairs. No constant curves were observed in either the human ratings or model-predicted CTI, so all curve-level correlations were well-defined.

In particular, the mean within-scene–product Spearman correlation of 0.8339 directly measures the agreement between model-predicted CTI and human click-intent rankings as target shares vary for the same products within the same scene.

For Weakness 2:

Section 4.3 says expected CTI is averaged over historical scenes, but does not explain whether q_{ik} is averaged by product identity, by product position within a scene (e.g., Product A vs. Product B), or over fixed product groups. Please provide the public q table used in Table 6 and explain how it is built

qikq_{ik} denotes the expected CTI for product slot i when generation uses the target-share vector specified by configuration k. We estimate it by averaging the predicted CTI scores of all products occupying slot i across historical scenes under configuration k. The resulting CTI table is not conditioned on product category or scene; instead, it maps target-share vectors to expected CTI vectors. For example, “Product A” denotes a product slot rather than a fixed product identity, and we average the CTI scores of all products occupying that slot across historical scenes.

In our paper, we construct a CTI table using 1,000 scenes and use this table to study the subsequent pricing mechanisms. The table is shown below:

Target shares (A, B)(0.95, 0.05)(0.90, 0.10)(0.85, 0.15)(0.80, 0.20)(0.75, 0.25)(0.70, 0.30)(0.65, 0.35)(0.60, 0.40)(0.55, 0.45)(0.50, 0.50)(0.45, 0.55)(0.40, 0.60)(0.35, 0.65)(0.30, 0.70)(0.25, 0.75)(0.20, 0.80)(0.15, 0.85)(0.10, 0.90)(0.05, 0.95)
Product A CTI0.93170.93170.93170.93140.92910.92260.91020.89100.87180.87090.72600.66610.56290.43840.33390.27110.24760.22590.2189
Product B CTI0.28220.28220.28730.32610.43390.54890.64770.71470.76150.82600.84880.86220.87450.89070.89110.89110.89110.89290.9024

As mentioned above, Product A and Product B in the table do not correspond to specific products. Instead, each entry represents the expected CTI across products occupying the corresponding slot under a given share configuration. The table thus reflects the expected CTI for each product slot in the generated image when a particular target-share vector is supplied to the generation model.

We additionally conduct an experiment comparing two CTI estimation methods:

  1. Averaging by product slot and target share, as adopted in our paper, which aggregates CTI scores across training scenes for the same prompt position (A or B) and target share, regardless of product identity.
  2. Averaging by product identity and target share, which aggregates historical CTI scores only for the same product identity at the same target share. Product identity is determined by the specific brand and product description.

For this experiment, we construct 100 two-advertiser scenes, generating 19 images per scene, one for each target-share configuration. The scenes cover 20 distinct products, with each product appearing in 10 scenes. This allows us to investigate whether averaging historical CTI by product identity improves CTI prediction accuracy in new scenes.

We use five-fold cross-validation. In each fold, we use 80 scenes to construct the CTI tables for both estimation methods and evaluate prediction errors on the remaining 20 held-out scenes. We compute MAE between the CTI estimated from the historical lookup table and the CTI predicted directly from the generated images in the held-out scenes.

CTI table estimation methodMAE ↓ (95% CI)
Averaging by product slot0.1674 [0.1597, 0.1753]
Averaging by product identity0.1590 [0.1516, 0.1666]

We further evaluate the two CTI estimation methods on our original dataset of 1,000 scenes. According to our statistics, each product appears in 3–6 scenes. We conduct five-fold cross-validation with scene-level splits, obtaining MAEs of 0.1752 [0.1689, 0.1824] for slot-level aggregation and 0.1810 [0.1761, 0.1883] for product-specific aggregation, respectively. In this setting, slot-level aggregation achieves better performance.

We believe that when sufficient historical observations are available for each product, product-specific CTI estimation can better preserve product-specific characteristics, leading to more accurate estimates. However, when a product appears only a limited number of times in historical data, its estimated CTI may be biased toward the particular scenes in which it appeared, potentially reducing its generalization accuracy in new scenes.

Considering that products participating in real-world auctions may not have sufficient historical observations, product-specific estimation cannot always rely on adequate historical data. Therefore, we adopt the simpler slot-level aggregation strategy, which estimates the average CTI that the generation model produces for each product slot under different target-share configurations.

We acknowledge that the slot-level CTI estimation adopted in our paper is a simplified approximation. In practical applications, platforms could develop more fine-grained CTI estimation models by incorporating richer product-level historical data, scene characteristics, and other relevant factors to improve estimation accuracy and practical utility. Such improvements are compatible with our proposed framework and would not require modifications to its core allocation and payment rules.

please also state how the 19 configurations are constructed in the three-advertiser setting.

For the three-advertiser setting, we use a representative discrete set of 19 paid configurations: the equal-share configuration (1/3,1/3,1/3)(1/3,1/3,1/3), together with all distinct permutations of (0.4,0.3,0.3)(0.4,0.3,0.3), (0.5,0.25,0.25)(0.5,0.25,0.25), (0.6,0.2,0.2)(0.6,0.2,0.2), (0.7,0.2,0.1)(0.7,0.2,0.1), and (0.8,0.1,0.1)(0.8,0.1,0.1), yielding 1+3+3+3+6+3=191+3+3+3+6+3=19 configurations in total. A separate no-allocation option is also included.

This construction includes equal allocation, varying degrees of exposure concentration among advertisers, and allocations in which all three advertisers receive different shares. It also provides broad coverage of share values between 0.1 and 0.8: 0.1,0.2,0.25,0.3,1/3,0.4,0.5,0.6,0.7,0.1, 0.2, 0.25, 0.3, 1/3, 0.4, 0.5, 0.6, 0.7, and 0.80.8.

Since exhaustively enumerating all possible allocations is infeasible, this set provides a finite discretization of the continuous share space.

For Question 1:

The tables report monotonicity only for the overall CTI score. How do the four individual dimensions behave as the target share increases

All four dimensions continue to exhibit strong monotonic trends, with each capturing a different aspect that influences CTI. The results are presented in the table below. This table provides some indication of how each dimension contributes to the monotonicity of the overall CTI score.

Scoring dimensionPearson ↑Spearman ↑Pairwise ↑Kendall’s τ-b ↑Isotonic R² ↑
Product Prominence0.75980.77810.76360.65960.8394
Brand Recognizability0.61710.63200.70270.52310.7067
Scene Integration0.60080.60910.68560.50570.7037
Contextual Value Communication0.64700.66750.70730.55860.7432

For Question 2:

Please describe how the curve-level metrics are aggregated in Tables 1 and 2 and in the appendix tables

For each product in each scene, we obtain a curve showing how CTI varies with the target share. We compute the five metrics separately for each curve and then average each metric across all curves.

how constant or degenerate curves are handled for metrics such as Spearman correlation and Isotonic R^2, and the uncertainty around these aggregated values.

For constant curves with identical CTI scores across all configurations, we assign a value of 0 to the otherwise undefined Pearson correlation, Spearman correlation, Kendall’s τb\tau_b, and Isotonic R2R^2. Pairwise monotonicity equals 0.5 by definition. These curves are included in the final arithmetic means.

To quantify the uncertainty in the aggregated estimates, we report the 95% confidence intervals below:

MetricMean95% Confidence Interval
Pearson0.7222[0.7146, 0.7349]
Spearman0.7370[0.7292, 0.7505]
Pairwise0.7603[0.7574, 0.7658]
Kendall’s τ-b0.6027[0.5959, 0.6153]
Isotonic R²0.8126[0.8044, 0.8264]

To evaluate the effectiveness of our method at the individual scene level, we conduct the following additional experiment.

For each product in each scene, we compare the metrics computed from the CTI–share curve produced by our method with those computed from the corresponding baseline curve. For each metric, we report the win rate, defined as the proportion of scene–product pairs for which our method achieves a higher value than the baseline. The results are shown in the table below:

MetricWin Rate
Pearson79.26%
Spearman80.18%
Pairwise75.58%
Kendall’s τ-b79.26%
Isotonic R²72.35%

This comparison shows that our method achieves better monotonicity for the vast majority of scene–product pairs.

Dear Reviewer ZZ1H,

Thank you again for your thoughtful review. We are following up on our responses and additional experiments, which address your concerns as follows:

  • W1: We clarified that the human–model correlations are computed within each scene–product pair and then averaged. The mean Spearman correlation of 0.8339 therefore measures ranking agreement across configurations for the same product in the same scene.
  • W2: We provided the public CTI table, explained its construction and the meaning of the advertiser indices, and specified all 19 three-advertiser configurations. We also added held-out experiments comparing aggregation by product slot and by product identity.
  • Q1 and Q2: We reported monotonicity for all four scoring dimensions, clarified metric aggregation and constant-curve handling, and provided 95% confidence intervals and scene–product-level comparisons with the baseline.

We would appreciate your feedback on these responses during the discussion period, particularly whether any concerns remain unresolved. We would be happy to provide further clarification. If the additional evidence addresses your concerns, we would also be grateful if you could reconsider your rating accordingly.

Thank you for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.