PaperlayerView exact revision on arXiv

Source-linked AI summary

DiffAuction: An Auction-Driven Framework for Generative Image Advertising

Yijie Su, Ruiyang Si, Tonghan Wang

arXiv:2610.34917v1cs.CV

TL;DR

Generative image advertising must control product exposure, estimate image-specific product CTI, and price jointly feasible allocations using advertisers’ private values. DiffAuction combines share-conditioned generation, CTI evaluation, and a truthful learnable auction; experiments report stronger share–CTI monotonicity and higher auction revenue than baselines.

  • Problem

    Shared-image advertising requires controlling exposure shares, estimating each generated image’s product-level CTI, and pricing jointly feasible configurations using advertisers’ private values.

    Source: §1 · pp. 1–2

  • Method

    DiffAuction combines RSC and SCP for share-controlled generation, VLM–LLM product-level CTI assessment, and a learnable affine-maximizer auction over feasible configurations.

    Source: §1 · p. 2

  • Results

    DiffAuction strengthens target-share–CTI monotonicity and reduces generation time, while its learned AMA achieves higher platform revenue than baselines in simulated-valuation experiments.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The framework connects controllable image exposure, interpretable product-level click-intent estimates, and truthful pricing with DSIC and IR guarantees.

    Source: §5 · p. 9

    The mechanism assumes the platform has access to the joint distribution of advertisers’ private valuations.

    Source: §3.3 · p. 6

Abstract

from arXiv · show

Diffusion models enable multiple advertisers to compete for exposure within a shared generated 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, a framework integrating 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. 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 1000 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.

1 INTRODUCTION

DiffAuction addresses the challenge of controlling product exposure in shared generated images, evaluating each product’s click intent, and auctioning feasible exposure configurations. It combines share-conditioned generation, a product-level CTI evaluator, and a revenue-oriented auction mechanism.

1 INTRODUCTION

  • The setting requires generation to realize auction-selected exposure shares, content-based estimates of each image’s product-level CTI, and payments that incentivize truthful reports.Each generated image is a new outcome, unlike a predefined ad slot with historical click-through rates.Source: §1 · pp. 1–2
  • RSC uses one joint model pass to calibrate product-specific attention residuals to target shares, while SCP adjusts the viewing perspective to support the desired allocation.Together, they address the computational cost of multiple conditional evaluations and the tendency of fixed prompts to leave products obscured or overly prominent.Source: §1 · p. 2
  • The VLM–LLM evaluator estimates product-level CTI from visual evidence across four dimensions and correlates strongly with judgments from ten annotators on 1,000 images.The dimensions are product prominence, brand recognizability, scene integration, and contextual value communication.Source: §1 · p. 2
  • The auction selects among jointly feasible generation configurations using expected CTI and configuration-specific private values, then sets advertiser-specific payments with DSIC and IR guarantees.The formulation allows advertisers’ values per CTI unit to vary across configurations.Source: §1 · p. 2
  • The combined RSC–SCP method reduces generation time and strengthens the relationship between target exposure shares and product-level CTI; the learned AMA earns higher revenue than baselines.The revenue comparison uses simulated private valuations.Source: §1 · p. 2

2 RELATED WORK

Prior work explores auctions over generated language and inference-time control of diffusion models, but generated-image advertising combines multiple products in one scene and targets their relative exposure.

2 RELATED WORK

  • LLM-native advertising auctions treat generated content as inventory, with mechanisms allocating influence over tokens, discourse segments, responses, or internal representations.Examples include token-level auctions (Duetting et al., 2024), RAG-based segment auctions (Hajiaghayi et al., 2024), MOSAIC (Soumalias et al., 2025), LLM-Auction (Zhao et al., 2025), and Neuron Auctions (Yun et al., 2026).Source: §2 · p. 3
  • Sun et al. (2025) condition generation on target shares but require multiple model evaluations per denoising step, use an unvalidated and less interpretable CLIP proxy, and omit configuration-specific valuations.Their work also does not unify image generation and auction design to optimize platform revenue.Source: §2 · p. 3
  • Unlike concept editing, multi-product image advertising regulates relative exposure in a shared scene; SCP and RSC support target-share conditioning with one forward pass per denoising timestep.Prior inference-time diffusion controls intervene through attention, guidance, or internal representations.Source: §2 · p. 3

3 METHOD

DiffAuction links target-share-controlled image generation to product-level CTI evaluation and an auction that selects a shared generation configuration and prices advertiser exposure. Its method combines residual calibration and share-conditioned prompting, a content-based CTI proxy, and a learnable affine maximizer.

3.1 RESIDUAL SHARE CALIBRATION AND SHARE-CONDITIONED PROMPTING

  • RSC decomposes joint-attention image-feature updates into product-specific residuals and rescales them to match target shares while preserving total residual magnitude and leaving unchanged terms intact.The stated guarantee requires each product with a positive target share to have a nonzero residual.Source: §3.1 · pp. 3–4
  • SCP maps each product’s target share relative to equal allocation to one of five ordered prominence levels, complementing RSC’s real-valued residual-norm calibration.Higher levels encourage favorable views and greater visibility; lower levels instruct the model to make the product less visually prominent.Source: §3.1 · pp. 4–5

3.2 EVALUATING PRODUCT-LEVEL CLICK-THROUGH INTENT

  • The CTI proxy scores product prominence, brand recognizability, scene integration, and contextual value communication using visual evidence extracted by a VLM and a fixed-rubric LLM.The equally weighted mean of dimension scores is rescaled to [0, 1].Source: §3.2 · p. 5

3.3 LEARNABLE AFFINE-MAXIMIZER AUCTION MECHANISM

  • The learnable affine maximizer selects one common configuration and sets advertiser-specific payments; its induced menus support truthful reporting and guarantee DSIC and IR.The platform learns advertiser weights and configuration offsets to maximize expected revenue under truthful reporting.Source: §3.3 · pp. 5–6
  • Each public configuration specifies target shares and expected CTI, while advertisers have private values per CTI unit that may vary across configurations.The configurations’ share and CTI vectors are fixed before bidding; configuration 0 represents no allocation.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

The experiments evaluate share–CTI control, CTI agreement with human judgments, and auction revenue. The learned AMA leads the revenue comparisons, while the generation method strengthens monotonicity and runs faster than the baseline.

4.1 MONOTONICITY BETWEEN TARGET EXPOSURE SHARES AND CTI

  • The method performs best across all five monotonicity metrics, and target-share increases yield a clear increasing trend in average CTI across two- and three-advertiser settings.The evaluation covers 1,000 two-product scenes, 400 products, and 19 share configurations per scene; Figure 2 aggregates 1,000 scenarios.Source: §4.1 · p. 7, Table 1 · p. 7, Figure 2 · p. 7
  • RSC and SCP together achieve the best performance across all ablation metrics, and their gains over the same-SCP baseline cannot be attributed to prompting alone.The ablation covers 200 scenes and 19 share configurations ranging from (0.95, 0.05) to (0.05, 0.95).Source: §4.1 · p. 7, Table 2 · p. 8
  • The method is consistently faster per denoising timestep than Sun et al. (2025), with its advantage growing as product count increases.Inference times use 200 prompts on one NVIDIA A800 GPU and report mean ± standard deviation over ten runs.Source: §4.1 · p. 8, Table 3 · p. 8

4.2 VALIDATION OF THE CTI EVALUATOR

  • The VLM–LLM evaluator agrees with human judgments more than direct VLM scoring across all three metrics on 1,000 generated images.Ten trained annotators each rated 100 images using the same four-dimensional rubric.Source: §4.2 · p. 8, Table 4 · p. 8
  • All inter-annotator metrics have ordinal α > 0.8; CTI’s exact agreement is 42.0%, while its pairwise MAE is 0.0450.Exact agreement requires all five annotators to assign identical scores; the analysis covers 100 images from 20 scenes.Source: §4.2 · p. 8, Table 5 · p. 9

4.3 EVALUATING THE ECONOMIC PERFORMANCE OF THE AUCTION MECHANISM

  • 5.10% and 4.27% under correlated valuations, and 51.12% and 50.66% under independent valuations, are the learned AMA’s revenue gains over the Myerson-based baseline for two and three advertisers, respectively.It achieves the highest revenue in all four settings and consistently outperforms VCG.Source: §4.3 · p. 9, Table 6 · p. 9

5 CONCLUSION

DiffAuction combines controllable image generation, product-level CTI evaluation, and truthful auction design. Its experiments report stronger share–CTI monotonicity, greater generation efficiency, human agreement, and higher auction revenue.

5 CONCLUSION

  • RSC and SCP provide continuous exposure control with one model forward pass per denoising step.They are components of DiffAuction’s controllable image-generation framework.Source: §5 · p. 9
  • The CTI evaluator supplies interpretable product-level click-intent estimates for the framework.The conclusion describes these estimates as part of DiffAuction’s integrated approach.Source: §5 · p. 9
  • The learnable affine maximizer selects feasible configurations and payments using expected CTI and advertisers’ private values, while ensuring DSIC and IR.Experiments report improved share–CTI monotonicity, generation efficiency, human agreement, and auction revenue.Source: §5 · p. 9

A FURTHER EXPERIMENTS ON THE MONOTONICITY BETWEEN TARGET EXPOSURE SHARE AND CTI

Further experiments test share–CTI monotonicity with three advertisers, with larger numbers of products, and among products in the same category. The reported three-advertiser and larger-product results maintain the monotonicity pattern.

A FURTHER EXPERIMENTS ON THE MONOTONICITY BETWEEN TARGET EXPOSURE SHARE AND CTI

B.1 FURTHER EXPERIMENTS ON RSC

The RSC ablations examine how control duration, intervention location, and residual reweighting affect share–CTI monotonicity. Eight controlled steps and applying RSC to both transformer-stream types perform best in their respective comparisons.

B.1 FURTHER EXPERIMENTS ON RSC

  • Eight controlled denoising steps perform best: extending RSC to all ten slightly degrades the five monotonicity metrics, while two-to-eight steps improve them consistently.The authors attribute the ten-step decline to possible over-constraint during final refinement and identify eight steps as the best balance between monotonicity and overhead.Source: §B.1 · p. 14, Table 10 · p. 14
  • Applying RSC to both double-stream and single-stream blocks achieves the best performance across all share–CTI monotonicity metrics on 200 scenes.The authors explain that the stages support early product-token/image-feature interaction and subsequent multimodal representation refinement.Source: §B.1 · p. 14, Table 11 · p. 14
  • RSC outperforms direct residual scaling because it matches residual-norm shares to target shares while preserving the sum of product-specific residual norms.Direct scaling uses eRi = siRi; RSC uses eRi = (si/pi)Ri, where pi is each product’s fraction of summed residual norms.Source: §B.1 · pp. 14–15, Table 12 · p. 15

B.2 DEFINITIONS OF CORRELATION AND MONOTONICITY METRICS

This section defines metrics for assessing how target exposure share relates to CTI, then describes share-conditioned prompts and visualizes exposure-share control. The prompts map relative shares to prominence instructions that complement residual control.

B.2 DEFINITIONS OF CORRELATION AND MONOTONICITY METRICS

  • Pearson correlation measures linear association, Spearman correlation measures rank consistency including nonlinear relationships, and isotonic R2 measures fit to a nondecreasing CTI curve.A nonconstant nondecreasing curve, including one with plateaus, scores one on isotonic R2, so that score alone does not establish fine-grained share responsiveness.Source: §B.2 · pp. 15–16
  • Pairwise monotonicity scores higher-share/higher-CTI comparisons, while Kendall’s τb measures monotonic association while accounting for CTI ties.Pairwise scores assign one to an increase, one-half to a tie, and zero to a decrease; a constant CTI curve receives 0.5.Source: §B.2 · p. 15

B.3 SHARE-CONDITIONED PROMPTS

  • Share-conditioned prominence computes each product’s target share relative to equal allocation (ri = 1), then maps positive relative shares to five levels; α = 0.5 and δ = 0.2 yield boundaries 0.5, 0.8, 1.2, and 1.5.Source: §B.3 · p. 16
  • These prompts encourage favorable views for higher-share products and complement residual control by allowing viewpoint and layout changes under a fixed prompt.The section also reports an analysis of α and δ across three parameter configurations on 50 scenes, but the supplied passages do not state its results.Source: §B.3 · p. 16

B.4 VISUALIZATION OF EXPOSURE-SHARE CONTROL

  • As target share increases, the product’s exposure in the generated image also increases across five scenes and five evenly spaced configurations per scene.Figure 4 visualizes exposure-share control.Source: §B.4 · p. 17, Figure 4 · p. 17

C FURTHER EXPERIMENTS ON CTI EVALUATION

The section tests whether model-predicted CTI tracks human click intent, while the following auction-details subsection describes training and evaluation of the learned mechanism. Model scores show strong associations with independent human ratings.

C FURTHER EXPERIMENTS ON CTI EVALUATION

  • Five additional annotators rated 570 images from 30 scenes without access to the four-dimensional rubric, and their mean ratings were compared with model-predicted CTI.Each scene contributed 19 images, and annotators independently rated the shuffled complete set.Source: C FURTHER EXPERIMENTS ON CTI EVALUATION · p. 17
  • Model-predicted CTI correlates with mean human click-intent ratings at Pearson 0.8526 and Spearman 0.8339, with ICC(2,1) of 0.7694.The authors interpret these results as further support for CTI as a proxy for average self-reported click intent within the evaluated sample.Source: C FURTHER EXPERIMENTS ON CTI EVALUATION · p. 17

D ADDITIONAL DETAILS ON THE AUCTION MECHANISM

  • The auction details train a separate AMA for each advertiser-count and valuation-distribution setting, select models by validation revenue, and test on independent profiles.The selected model is evaluated on 131,072 test profiles; reported revenues use payments for selected configurations rather than the training softmax approximation.Source: D ADDITIONAL DETAILS ON THE AUCTION… · p. 18

E PROOFS OF DSIC AND IR

The proof section establishes DSIC and IR for the affine maximizer auction, then derives truthfulness from report-independent prices and common menu optimality. Its guarantees cover any finite configuration set and advertiser count.

E PROOFS OF DSIC AND IR

  • The affine maximizer allocation and payment rules satisfy DSIC and IR for any number of advertisers and any finite configuration set.The stated setting fixes the public configuration set, expected-CTI table, and mechanism parameters independently of current reports.Source: E PROOFS OF DSIC AND IR · p. 18
  • The proof permits configuration-specific private values that need not be equal, independent, or monotone, and requires no distributional assumption on private types.Advertiser value for a configuration is modeled as expected CTI times its private value per CTI unit.Source: E PROOFS OF DSIC AND IR · p. 18

E.1 REPORT-INDEPENDENT MENUS AND COMMON OPTIMALITY

  • Each advertiser’s price menu is independent of its own report, and shared tie-breaking makes every advertiser select the same utility-maximizing configuration.The proof uses a positive advertiser weight and a price that remains independent of that advertiser’s report.Source: §E.1 · pp. 18–19

E.2 DOMINANT-STRATEGY INCENTIVE COMPATIBILITY

  • Truthful reporting is weakly dominant because a deviation can change the selected configuration but cannot change the advertiser’s menu prices.The argument allows simultaneous changes to multiple report coordinates, does not require other advertisers to report truthfully, and treats ties as yielding equal true utility.Source: §E.2 · p. 19

E.3 INDIVIDUAL RATIONALITY

The mechanism guarantees nonnegative truthful utility and nonnegative payments, with zero payments when configuration 0 is selected. These guarantees have stated deployment and valuation-model conditions.

E.3 INDIVIDUAL RATIONALITY

  • Truthful participation yields nonnegative utility for every advertiser with a nonnegative true type and any profile of other reports.The zero-priced menu entry may differ by advertiser and need not be configuration 0.Source: §E.3 · p. 19

E.4 NONNEGATIVE PAYMENTS AND NO TRANSACTION

  • All menu prices and realized payments are nonnegative, and selecting configuration 0 makes every advertiser’s realized payment zero.The proof also states that the selected-entry payment is nonnegative.Source: §E.4 · p. 20
  • The guarantees hold for fixed strictly positive advertiser weights and deployed hard-max allocation, without requiring optimal revenue training, independent private types, or monotone CTI.They are conditional on V_i(k) = q_ikv_ik; expected DSIC and IR follow if this equals true expected image value, which a fixed historical CTI estimate alone does not establish.Source: §E.4 · p. 20

F THEORETICAL BASIS OF PLACEMENT-SPECIFIC EVALUATION DIMENSIONS

This section grounds four placement-specific evaluation dimensions in a two-axis framework and presents prompts for separating visual-evidence extraction from evidence-based CTI scoring. The scoring prompt applies 1–5 rubrics to prominence, brand recognizability, scene integration, and contextual value communication.

F THEORETICAL BASIS OF PLACEMENT-SPECIFIC EVALUATION DIMENSIONS

  • The evaluation scope prioritizes placement-relevant factors reliably assessable from generated images, excluding subjective or abstract judgments such as overall aesthetics.The four metrics are not claimed to exhaust CTI determinants or every property in their conceptual cells; consumer-specific factors such as purchase intention and relevance to a particular user are outside scope.Source: F THEORETICAL BASIS OF… · pp. 20, 22, 23
  • The framework crosses Stimulus (product versus product–scene relation) with Processing (explicit versus implicit) to form four conceptual cells.Stimulus distinguishes the inserted product from its relation to the scene; Processing distinguishes directly observable configurations from interpretations of higher-level commercial meaning.Source: F THEORETICAL BASIS OF… · p. 21
  • The four cells operationalize Product Prominence, Brand Recognizability, Scene Integration, and Contextual Value Communication, covering visual presence, brand meaning, scene compatibility, and contextual commercial meaning.These correspond respectively to Product×Explicit, Product×Implicit, Product–Scene Relation×Explicit, and Product–Scene Relation×Implicit.Source: F THEORETICAL BASIS OF… · pp. 21–22
  • The examined in-scope alternatives substantially overlap the four metrics, while others fall outside the study’s scope, supporting their adequacy for placement evaluation—not completeness for all CTI determinants.The scope excludes criteria that cannot be reliably judged from image content or are designed for conventional advertisements rather than AI-generated product placements.Source: F THEORETICAL BASIS OF… · p. 22, Table 15 · p. 22

G.1 VISUAL EVIDENCE EXTRACTION PROMPT

  • The VLM extraction prompt requests product-specific evidence across the four dimensions, prioritizing visible facts and requiring uncertainty rather than fabricated details or effectiveness judgments.It distinguishes product visibility from brand confirmation and asks for evidence on prominence, brand cues, physical scene integration, and contextual value communication.Source: G PROMPTS FOR CTI EVALUATION · p. 24, §G.1 · pp. 24–25

G.2 CTI SCORING PROMPT

  • The LLM scoring prompt assigns each product an integer score from 1 to 5 on the four dimensions using only the VLM-provided evidence, with dimensions evaluated independently.The task is single-image, prohibits unsupported information or comparisons across images, and bars high scores when the target product is not visible.Source: §G.2 · pp. 25–26
  • The prompt scores visual salience, specific-brand identification, natural scene integration, and communication of product use or value on four 1–5 dimensions.It warns against scoring based merely on category-similar products, other brands’ logos, or textual product mentions.Source: §G.2 · pp. 26–28

Related papers

Loading 2610.34917v1…