Trait-Targeted Ad Generation

• Trait-targeted ad generation is the automated creation of personalized ads by leveraging user traits, psychometrics, and contextual signals.
• It utilizes diverse methodologies including NLG with control codes and diffusion models, integrating trait detection with reward-based creative optimization.
• Empirical results demonstrate significant uplifts in CTR, engagement, and revenue through methods like determinantal sampling and probabilistic profiling.

Trait-targeted ad generation refers to the automated creation of advertising content—textual, visual, or multimodal—conditioned on user traits, interests, personality dimensions, context, or fine-grained categories. The objective is to optimize ad relevance and efficacy by aligning creative elements with the inferred or measured characteristics of individual users or defined cohorts. This paradigm spans natural language generation (NLG) systems with category or personality control, generative diffusion pipelines for creative images, and large-scale semi-supervised modeling of user traits to drive both creative design and targeting. The following sections present the major architectures, modeling approaches, conditioning mechanisms, and empirical findings in trait-targeted ad generation.

1. Trait and Category Representation

Trait-targeted ad generation relies on explicit or implicit user segmentation, leveraging traits such as gender, interests, and psychometric dimensions, or product-centric categories.

• Personality and Psychometrics: Experiments with GPT-3.5 for text ad generation have directly targeted Big Five personality traits (notably Openness and Neuroticism), using SAPA Big Five questionnaire scores to label participants as “high” or “low” on each trait (Meguellati et al., 3 Dec 2025).
• Category Conditioning: Systems like DeepGen structure ad generation around a fixed set $C\approx 12$ of categories or “control codes” (e.g., PRODUCT, BRAND, DISCOUNT) that correspond to key selling points (Golobokov et al., 2022).
• Demographics and Interests: Tumblr’s production targeting leveraged gender and a two-level interest taxonomy for large-scale sponsored post personalization, using editorially-labeled seeds to bootstrap a semi-supervised embedding model over billions of user-generated content items (Grbovic et al., 2016).
• Visual Context and Preferences: AdBooster and CG4CTR both condition visual creatives on text prompts representing inferred user interests, context (e.g., “Seasons”, “Events”, “Sports”), or group segmentation (e.g., demographic buckets) (Shilova et al., 2023, Yang et al., 2024).

These representations are typically either computed directly from explicit user data (such as declared gender or trait test results), constructed by editorial labeling and embedding expansion, or inferred through behavioral data and deep learning–based classifiers.

2. Conditioning Mechanisms and Model Architectures

Trait conditioning is implemented via several mechanisms:

• Control Codes for Language Generation: In DeepGen, a pretrained controllable NLG model (e.g., UniLMv2 or BART) is trained with each input prepended by a control code specifying the desired category or orientation. During multi-code inference, all codes are iteratively prepended to generate diversified, trait-targeted ad assets (Golobokov et al., 2022).
• Prompt Engineering for LLMs: LLM-based approaches such as those in (Meguellati et al., 3 Dec 2025) prompt GPT-3.5 using minimal templates reflecting desired trait targets: e.g., “Write a one-line ad … targeting people high in openness, without mentioning the trait explicitly”.
• Trait-Conditioned Prompts for Diffusion Models: Both AdBooster and CG4CTR supply trait- or context-conditioned text prompts (such as directly concatenated “query for context” phrases, or GPT-3.5–generated scene descriptions) to the text encoder of a diffusion-based image generator (Shilova et al., 2023, Yang et al., 2024).
• Deep Embedding Fusion: In image pipelines, conditioned embeddings are injected as additional channels or through cross-attention within diffusion architectures (notably all cross-attn blocks in the UNet backbone), modulating the denoising process to reflect user traits (Shilova et al., 2023, Yang et al., 2024).
• Probabilistic User Profiling: Tumblr’s system computes continuous-valued interest and gender scores per user, updating them nightly via time-decayed sums over new user-generated content and mapping keywords to taxonomy categories through a semi-supervised skip-gram model (Grbovic et al., 2016).

A key feature is the intentional separation of trait definition, trait detection or labeling, and its operationalization as a control signal in the ad creative pipeline.

3. Pipeline Designs for Trait-Targeted Generation

Linguistic Systems: DeepGen

DeepGen partitions ad generation into two main phases: (1) offline asset generation and selection, and (2) online, real-time asset selection and assembly (Golobokov et al., 2022).

• Offline Phase: Web pages are crawled, document understanding is applied, and a model ensemble generates diversified textual assets (titles, descriptions) per URL and control code. These assets are filtered for factuality and then embedded via a CDSSM model. Diversity is maximized by selecting a subset via determinantal point process (k-DPP) MAP inference.
• Online Phase: For each auction (user query), pre-computed assets are retrieved and scored per ad slot using logistic regression models. A contextual Thompson sampling bandit framework is used for asset selection, supporting both exploitation of high-performing fragments and exploration for diversity and freshness. End-to-end latency is <5 ms.

Visual Content: AdBooster and CG4CTR

Both AdBooster and CG4CTR combine Stable Diffusion–based image outpainting/inpainting with explicit trait- or group-conditioned prompting (Shilova et al., 2023, Yang et al., 2024).

• Creative Generation: Provided with a product image and binary mask preserving the object, these models generate diverse backgrounds or settings through text prompts incorporating user or cohort information.
• Reward-Based Optimization: CG4CTR integrates a reward model that fuses visual, textual, and caption features to predict CTR, serving as both a selector/ranker and a training signal in their self-cyclic alternated optimization loop.
• Prompt Selection: Transformer-based prompt models map user traits and product metadata to a distribution over prompt tokens, selecting those maximizing predicted CTR for each user group; prompt optimization and LoRA-based diffusion fine-tuning are alternated for stability.

The key technical advance in both architectures is the unification of trait-informed prompt selection, creative generation, and downstream reward-driven selection in a closed loop.

4. Evaluation Methodologies and Empirical Results

Trait-targeted ad generation efficacy is quantified through both offline metrics (e.g., cosine similarity of user and creative in embedding space, FID scores, MSE on predicted vs. realized CTR) and online business metrics (CTR uplift, RPM, user engagement).

Textual Systems

• DeepGen: Deployed at Bing, serving ~4% of global search ads. In large-scale A/B tests, DeepGen increased Revenue-per-Mille by 24.9%, CTR by 13.3%, Impression Yield by 11.9%, and Quick-Back-Rate by 5.3% over the best extraction baseline (Golobokov et al., 2022).
• LLM-Based Trait Ads: Targeting Openness, GPT-3.5–generated one-liners matched or slightly exceeded human-written ads on Product Rating (4.14 vs. 3.71, p=0.02), Purchase Intention (4.14 vs. 3.69, p=0.02), and Engagement Intention (4.33 vs. 3.73, p=0.01). Aggregate preference for AI ads was 51.1% vs. 48.9% for human-crafted, with no significant difference in trait-matched scenarios (Meguellati et al., 3 Dec 2025).

Visual Systems

• AdBooster: Generated creatives (outpainted with user-context-conditioned prompts) improved CLIP-based trait alignment scores by 10–21% over strong static baselines across various context clusters (seasons, events, sports). FID improved by ~16% after fine-tuning (Shilova et al., 2023).
• CG4CTR: Online uplift from trait-targeted visual creatives reached up to +10.4% CTR and +9.7% revenue over seller-supplied visuals after 10 self-cyclic iterations. Removing trait conditioning from the prompt model diminished uplift by ~3 percentage points. Ablations confirmed the importance of trait-aware prompts and Transformer fusion in the reward model (Yang et al., 2024).

Large-Scale Social Systems

• Tumblr: Gender and interest targeting based on semi-supervised neural embeddings drove a 20% lift in engagement over untargeted sponsored posts, with substantially higher lifts for some categories (e.g., +43% for Style & Fashion). The reach exceeded 90% of daily active users (Grbovic et al., 2016).

5. Practical Considerations and Deployment Lessons

Trait-targeted ad generation at web scale raises distinct challenges:

• Factuality Control: NLG models can hallucinate brand or offer details. Production deployments (e.g., DeepGen) couple train- and infer-time filtering on blacklists, brand linkages, and domain validity, raising factuality rates from ~91% to ~97% (Golobokov et al., 2022).
• Latency and Throughput: Optimizations such as FastSeq, batch-level n-gram blocking, and parallelization allow ≳200 URL/s per GPU and <5 ms per query for real-time assembly (Golobokov et al., 2022).
• Automated Data Augmentation: Absence of large labeled datasets for visual creative outpainting is mitigated by automated background masking, synthetic prompt generation, and filtering rules yielding tens of thousands of training examples for diffusion models (Shilova et al., 2023).
• Self-Cyclic Training: Alternated optimization of prompt and generation/LoRA weights, guided by a fixed reward model, stabilizes deep multimodal pipelines and prevents catastrophic forgetting. This regime is empirically superior to joint or naive sequential updates (Yang et al., 2024).
• Scalability: Nightly batch pipelines using MapReduce (Tumblr) and stateless web-scale serving orchestrate trait updates and ad selection for >80 million active users with linear complexity (Grbovic et al., 2016).

A summary table of key empirical results:

System	Main Trait Target(s)	Main Metric (Uplift/Result)	Reference
DeepGen (Bing)	Ad category (12 codes)	+24.9% RPM, +13.3% CTR over baseline	(Golobokov et al., 2022)
GPT-3.5 LLM study	Openness, Neuroticism	Parity with human, 4.14 vs 3.71 rating (O)	(Meguellati et al., 3 Dec 2025)
AdBooster	User interests/context	+10–21% CLIP sim, 16% FID improvement	(Shilova et al., 2023)
CG4CTR	User group, demographics	+10.4% CTR, +9.7% revenue (10 cycles)	(Yang et al., 2024)
Tumblr system	Gender, interest category	+20% engagement, >90% user coverage	(Grbovic et al., 2016)

6. Research Directions and Limitations

Key challenges and future research avenues include:

• Fine-Grained Trait Conditioning: While current systems effectively model demographic and categorical traits, extension to more nuanced psychographic, purchasing, and contextual signals (e.g., learned embeddings instead of text-based prompts) remains an open area (Shilova et al., 2023, Yang et al., 2024).
• Full-Creative Generation: Most visual pipelines focus on background/presentation adjustment; integrating text overlays and layout optimization remains underexplored in large-scale deployments (Yang et al., 2024).
• Human Evaluation: Although CLIP and FID metrics correlate with alignment, comprehensive human preference studies for visual trait targeting are mostly lacking in current diffusion-based literature (Shilova et al., 2023).
• Generative Model Capacity: The use of more expressive backbones (e.g., SDXL, ControlNet, Composer) and mixture-of-experts architectures could yield greater trait/personality fidelity and creative diversity (Shilova et al., 2023).
• Optimization Integration: Jointly optimizing the item ranking (recommendation) and creative selection/generation stages remains an open opportunity in the quest for end-to-end trait-aligned conversion maximization (Yang et al., 2024).

7. Generalization and Theoretical Insights

The current state of trait-targeted ad generation demonstrates the feasibility and scalability of end-to-end systems that integrate trait representation, controllable generative modeling, reward-based optimization, and real-time delivery. Frameworks combining supervised seed-labeled traits, semi-supervised representation learning, and explicit control signals are structurally general: any trait with a modest lexicon or labeling seed and observable behavioral correlates can be injected into such pipelines (Grbovic et al., 2016). Theoretical advances in determinantal point processes (diversity sampling) and context-conditioned bandits further enhance both the diversity and competitive performance of generated ads (Golobokov et al., 2022).

Empirically, LLM and diffusion-based models have achieved parity or superiority relative to expert-crafted baselines in both text and visual domains, with significant operational improvements in engagement and monetization metrics (Golobokov et al., 2022, Meguellati et al., 3 Dec 2025, Shilova et al., 2023, Yang et al., 2024, Grbovic et al., 2016). These systems underpin a new standard for scalable, data-driven, and trait-personalized creative optimization in digital advertising.