DiffAuction: An Auction-Driven Framework for Generative Image Advertising

Yijie Su; Ruiyang Si; Tonghan Wang

cs.CVOct 26, 2026

DiffAuction: An Auction-Driven Framework for Generative Image Advertising

Authors: Yijie Su, Ruiyang Si, Tonghan Wang

Organizations: Tongji University · Tsinghua University · Shanghai Qi Zhi Institute

Abstract

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.

Figures & tables

Figure 1: Overview of DiffAuction. DiffAuction connects generation, CTI evaluation, and auction. Given target shares, RSC and SCP control the exposure of products during generation. The resulting images are evaluated offline by a VLM–LLM pipeline that estimates product-level CTI. During the auction, a learnable affine maximizer combines the expected CTI of each configuration with advertisers’ reported values per CTI unit to select a shared configuration and determine advertiser-specific payments. The selected exposure shares are used to generate the advertisement image.

Method	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
FLUX (Prompt-only)	0.2181	0.2344	0.5727	0.1837	0.2802
Sun et al. (2025)	0.6675	0.6630	0.7343	0.5332	0.7708
Ours	0.7222	0.7370	0.7603	0.6027	0.8126

Table 1: Monotonicity between target exposure shares and CTI.

Panel of Figure 2 — Figure 2: Monotonicity between target share and CTI achieved by our method.

Method	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
SCP only	0.3039	0.3172	0.5989	0.2430	0.4897
RSC only	0.3730	0.3774	0.6236	0.2990	0.4964
Baseline + SCP	0.7651	0.7837	0.7797	0.6430	0.8880
RSC + SCP	0.8120	0.8346	0.8024	0.6981	0.9265

Table 2: Ablation study of the monotonic relationship between target exposure shares and product-level CTI across different method configurations, evaluated on 200 scenes.

# Products	Ours (s) $\downarrow$	Sun et al. (2025) (s) $\downarrow$	Speedup $\uparrow$	Time reduction $\uparrow$
2	$0.6654\pm 0.0002$	$1.2734\pm 0.0006$	$1.91\times$	47.74%
4	$0.7317\pm 0.0002$	$1.6992\pm 0.0004$	$2.32\times$	56.94%
6	$0.7985\pm 0.0018$	$2.5485\pm 0.0001$	$3.19\times$	68.67%
8	$0.8623\pm 0.0003$	$3.3980\pm 0.0017$	$3.94\times$	74.62%
10	$0.9289\pm 0.0026$	$4.2571\pm 0.0012$	$4.58\times$	78.18%

Table 3: Per-timestep inference-time comparison.

Evaluator	Pearson $\uparrow$	Spearman $\uparrow$	ICC(2,1) $\uparrow$
VLM only	0.662	0.574	0.610
VLM + LLM (ours)	0.669	0.603	0.653

Table 4: Agreement between predicted CTI and human judgments on 1,000 images from 200 scenes. Results are averaged over the two products. Higher values are better.

Measure	PP	BR	SI	CVC	CTI
Ordinal $\alpha$ $\uparrow$	0.8181	0.9654	0.8342	0.8419	0.9179
Exact (%) $\uparrow$	78.0	82.0	66.0	68.0	42.0
MAE $\downarrow$	0.1600	0.1400	0.2933	0.2733	0.0450

Table 5: Inter-annotator agreement among 20 scenes. Each image was independently rated by five annotators. We calculated agreement among the annotators’ ratings using all 100 images from the 20 scenes. PP, BR, SI, and CVC denote four dimensions.

Advertisers	Distribution	VCG	Myerson-based	AMA
2	Correlated	0.1592	0.4036	0.4242
2	Independent	0.3699	0.5648	0.8535
3	Correlated	0.2019	0.5238	0.5462
3	Independent	0.3498	0.6031	0.9086

Table 6: Mean platform revenue under truthful reporting.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
Sun et al. (2025)	0.5861	0.6579	0.7530	0.5449	0.6143
Ours	0.6839	0.7932	0.8171	0.6685	0.6922

Appendix

Table 7: Monotonicity between target exposure shares and CTI in the three-advertiser setting.

Setting	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
5 products	0.7084	0.7324	0.7773	0.6037	0.6992
10 products	0.5928	0.6004	0.7162	0.4981	0.5689

Appendix

Table 8: Monotonicity under increasing numbers of advertised products.

Figure 3: Monotonicity between target exposure share and CTI under extreme numbers of competing products.

Method	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
Sun et al. (2025)	0.4794	0.4858	0.6597	0.4007	0.6687
Ours	0.5600	0.5961	0.7073	0.4964	0.6755

Appendix

Table 9: Share–CTI monotonicity for competing products.

Control Steps	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$	Runtime (s) $\downarrow$
2	0.7581	0.7839	0.7788	0.6413	0.8604	$\mathbf{5.448\pm 0.003}$
4	0.7875	0.7925	0.7823	0.6522	0.8972	$6.422\pm 0.002$
6	0.7892	0.8074	0.7920	0.6717	0.8979	$7.401\pm 0.002$
8	0.8120	0.8346	0.8024	0.6981	0.9265	$8.378\pm 0.006$
10	0.8083	0.8249	0.7987	0.6893	0.9184	$9.344\pm 0.001$

Appendix

Table 10: Effect of the number of controlled denoising steps on share–CTI monotonicity and inference time. Runtime is averaged over three runs on an NVIDIA A800 GPU.

Controlled Blocks	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
Double-stream only	0.6878	0.7021	0.7444	0.5649	0.7861
Single-stream only	0.6478	0.6605	0.7297	0.5291	0.7588
Double- and single-stream	0.8120	0.8346	0.8024	0.6981	0.9265

Appendix

Table 11: Effect of applying residual control to different FLUX transformer streams.

Method	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
Direct scaling	0.7792	0.7705	0.7687	0.6232	0.8393
RSC	0.8120	0.8346	0.8024	0.6981	0.9265

Appendix

Table 12: Comparison of direct residual scaling and RSC.

Advertisers	$\alpha$	$\delta$	Pearson $\uparrow$	Spearman $\uparrow$	Pairwise $\uparrow$	Kendall’s $\tau_{b}$ $\uparrow$	Isotonic $R^{2}$ $\uparrow$
2	0.40	0.15	0.7716	0.7750	0.7802	0.6431	0.8788
2	0.50	0.20	0.7747	0.7829	0.7826	0.6462	0.8840
2	0.60	0.25	0.7626	0.7706	0.7758	0.6333	0.8886
3	0.40	0.15	0.6899	0.7590	0.7956	0.6370	0.6922
3	0.50	0.20	0.7042	0.7817	0.8071	0.6579	0.7048
3	0.60	0.25	0.6783	0.7560	0.7920	0.6340	0.6760

Appendix

Table 13: Hyperparameter search for threshold values.

Figure 4: Visualization of exposure-share control.

Stimulus	Processing
	Explicit	Implicit
Product	Product Prominence	Brand Recognizability
Product–Scene Relation	Scene Integration	Contextual Value Communication

Appendix

Table 14: The

2\times 2

conceptual organization of placement-specific visual evaluation.

Candidate criterion	Case-based analysis	Relation to framework
Attention Retention	Sustained attention is a temporal behavioral response that cannot be measured directly from the image alone.	Out of scope

Product Recognizability	Identifying a product category or specific item is related to identifying its brand.	BR

Viewpoint Favorability	Viewpoint affects product visibility, the visibility of brand cues, and the plausibility of the product’s presentation in the scene.	PP + BR + SI

Appendix

Table 15: Selected alternative evaluation criteria and their relationships with the proposed metrics.

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 463 papers

Oct 10, 2026

Diffusion-Based Auctions for Multi-Winner Banner Ads

We propose a new framework for banner ad auctions, leveraging diffusion models to allow for multiple bidders to co-win a single banner ad opportunity, and in particular influence the ad shown on the banner depending on the strength of their bids, enabling a far richer outcome space compared to standard banner ad auctions, where there is a single winner who shows their ad. In our framework, the outcomes are images (what ad to show) and each bidder expresses their preferences over outcomes by declaring their value per unit of alignment between the image chosen by the mechanism and a textual description of their desired image. Our proposed mechanism combines the advertisers' bids and textual descriptions into a score that drives the reverse process of a diffusion model. This is used to sample a set of candidate images from which the welfare maximizing one is eventually chosen. Extensive experimentation evaluates our proposed mechanism around the following axes: (1) efficiency improvements of our allocation rule relative to the single-winner allocation rule; (2) monotonicity of our allocation rule; and (3) approximate incentive compatibility when combining our allocation rule with Myerson lemma-based or Clarke rule-based payments. Improvements along all these axes are shown with the number of Monte Carlo samples. Meanwhile, these benefits are attained while preserving the image quality.

Lilian Sun, Henry Huang, Fucheng Warren Zhu +2

Harvard · MIT

Online Advertising Mechanism Design

Ask PDF

Jul 30, 2026

Evaluating and Pricing Advertisements in AI-Generated Responses

As search increasingly shifts toward LLM-driven answer engines, advertising is becoming embedded within the generated response itself and should therefore be evaluated for both user utility and commercial value. The key challenge is click-through intent: behavioural logs are unavailable, human annotation resists calibration, and frontier LLM judges conflate intent with linguistic fluency. These gaps compound, as principled pricing presupposes a continuous intent signal, while generating such a signal presupposes supervision that is currently unavailable. We construct the missing supervision through a psychologically grounded agent simulation framework, and distil it into a parameter-efficient evaluator that predicts click-through intent, together with the three companion dimensions of ad quality, as smooth, differentiable estimates. Validated through sign-certain behavioural perturbations, the evaluator surpasses frontier zero-shot judges on relevance sensitivity (79% versus 60-67%), tracks graded content degradation, generalises without error to 103 fictional products, and agrees with human preference in 86% of pairwise judgements across five annotators, with agreement rising in the evaluator's confidence. Upon its estimates we build the pricing layer directly, deriving the unique payment rule under which truthful bidding is optimal, demonstrating it on a best-of-k allocation, and extending the mechanism to non-monotone allocations. The same differentiable signal stands ready as a training objective for ad generation.

John L. Turner-Smith, Zimeng Huang, Yuhan Fu +2

Carnegie Mellon University, Pittsburgh, PA, USA · College of AI, Tsinghua University, Beijing, China

Mechanism Design Online Advertising Intent Classification CTR Prediction Automated Evaluation Generative AI Evaluation

Ask PDF

Aug 27, 2026

Token-Level Advertising

Generative AI is transforming how people access information, challenging traditional advertising mechanisms built around predefined slots. Towards generation-native advertising, we propose the Latent Advertiser Mixture Auction (LAMA), a token-level advertising mechanism that embeds advertiser influence directly into the generation process. Advertisers report local continuation values that induce advertiser-specific next-token policies, from which the platform decodes through a latent mixture while updating an allocation posterior. We show that LAMA satisfies Markov DSIC and IR, and achieves near-optimal KL-regularized welfare. We further develop a learning-based implementation that reconstructs the required reports online from learned local advantages and root values. Proof-of-concept experiments on real-world commercial-search query splits show that LAMA improves platform welfare and revenue while maintaining user-facing response quality, providing initial evidence for the feasibility of generation-native advertising.

Hanbing Liu, Bowei Zhang, Changyuan Yu +2

Renmin University of China Beijing, China · Beijing, China · Stanford University Stanford, CA, USA

Online Advertising Mechanism Design Game Theory

Ask PDF

Abstract

Image Generation

Explore similar work

Diffusion-Based Auctions for Multi-Winner Banner Ads

Evaluating and Pricing Advertisements in AI-Generated Responses

Token-Level Advertising