🤖 AI

One Size, Many Fits: Aligning Diverse Group-Wise Click Preferences in Large-Scale Advertising Image Generation

This paper introduces OSMF, a unified framework that addresses the limitations of one-size-fits-all advertising image generation by dynamically grouping users and employing a Group-aware Multimodal Large Language Model with Group-DPO fine-tuning to align diverse group-wise click preferences, thereby achieving state-of-the-art performance in both offline and online metrics.

Original authors: Shuo Lu, Haohan Wang, Wei Feng, Weizhen Wang, Shen Zhang, Yaoyu Li, Ao Ma, Zheng Zhang, Jingjing Lv, Junjie Shen, Ching Law, Bing Zhan, Yuan Xu, Huizai Yao, Yongcan Yu, Chenyang Si, Jian Liang

Published 2026-02-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuo Lu, Haohan Wang, Wei Feng, Weizhen Wang, Shen Zhang, Yaoyu Li, Ao Ma, Zheng Zhang, Jingjing Lv, Junjie Shen, Ching Law, Bing Zhan, Yuan Xu, Huizai Yao, Yongcan Yu, Chenyang Si, Jian Liang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are running a massive digital billboard that shows ads to millions of people. In the past, the strategy was "One Size Fits All." You would create one perfect-looking ad for a pair of shoes and show it to everyone. The goal was to get the highest total number of clicks.

But here's the problem: Not everyone likes the same thing.

A teenager might love a shoe ad with a gritty, urban skate-park background.
A grandparent might prefer the same shoe in a clean, sunny garden setting.
If you show the "skate-park" ad to the grandparent, they won't click. If you show the "garden" ad to the teen, they won't click either.

The paper introduces a new system called OSMF (One Size, Many Fits) to fix this. Instead of making one ad for everyone, it creates a unique ad for specific groups of people, all from the same computer brain.

Here is how it works, broken down into simple steps:

1. The "Smart Grouping" (PAAG)

The Analogy: Imagine a party planner who doesn't just sort guests by age or gender. Instead, they look at the menu (the product) and the guests together.

If the product is smartphones, the planner realizes that men and women might have very different tastes.
If the product is running shoes, the planner realizes that age matters more than gender.
What the paper does: The system uses a tool called PAAG (Product-Aware Adaptive Grouping). It looks at the product and the user's history to dynamically sort people into the right "clubs." It doesn't use rigid rules; it figures out, "For this specific product, these 50,000 people actually want the same thing."

2. The "Super-Translator" (G-MLLM)

The Analogy: Think of a master chef who can cook for a whole crowd, but usually just makes one big pot of soup. The new system is like a chef who can instantly switch recipes based on who is sitting at the table.

The system uses a Group-aware Multimodal Large Language Model (G-MLLM). This is a super-smart AI that understands both text and images.
Before it starts cooking, it is "pre-trained" to understand the specific "flavors" (preferences) of each group. It learns that Group A likes "bright colors" and Group B likes "minimalist styles."

3. The "Taste Test" (Group-DPO)

The Analogy: Imagine the chef makes two versions of a dish. A group of tasters (the reward model) tries them and says, "Group A loved the spicy one, but Group B hated it."

The system uses a technique called Group-DPO. It doesn't just ask, "Which image is better?" It asks, "Which image is better for this specific group?"
It compares two ads for the same product. If the "urban" ad gets more clicks from the teen group, the AI learns to make more "urban" ads for them. If the "garden" ad wins with the seniors, it learns that too.
This ensures the AI doesn't get confused by conflicting tastes. It learns to be a chameleon, changing its style for each group.

4. The "Massive Recipe Book" (GAIP Dataset)

The Analogy: To teach a chef how to cook for 40 million people, you need a massive cookbook with real feedback.

The researchers couldn't find a public book with this much data, so they wrote their own.
They created GAIP, a dataset containing 40 million users and 600,000 distinct groups. It's like having a record of exactly what 600,000 different types of people clicked on when they saw 40 million different ads. This is the first time such a huge, detailed "preference map" has been made public.

The Result

When they tested this system:

Offline (Simulations): It predicted clicks better than any previous method.
Online (Real World): When they actually showed the ads on a major e-commerce site, the "One Size, Many Fits" approach got significantly more clicks than the old "One Size Fits All" methods.

In short: The paper says, "Stop trying to please everyone with one ad. Use smart grouping to figure out who likes what, train a super-AI to understand those specific groups, and generate custom ads for each one. It works better, and we have the data to prove it."

Technical Summary: One Size, Many Fits (OSMF)

Problem Statement

Current large-scale advertising image generation systems predominantly rely on a "one-size-fits-all" strategy, optimizing for a single, aggregate Click-Through Rate (CTR) across the entire user base. While recent advancements have integrated CTR as a reward signal in Reinforcement Learning from Human Feedback (RLHF), this approach fails to account for the substantial heterogeneity of preferences among different user groups.

As noted in the introduction, treating overall user preferences as universally applicable leads to suboptimal performance for specific demographics. For instance, preferences for smartphone advertisements may differ significantly by gender, while shoe advertisements may vary more by age. A unified model attempting to satisfy these conflicting preferences (e.g., "natural scenes" vs. "sports field" backgrounds) simultaneously often results in images that fail to engage specific groups, thereby limiting targeted marketing effectiveness and wasting advertising spend. Furthermore, existing group-wise RLHF approaches face two critical challenges: (1) Limited Scalability of User Grouping, where static, fixed-attribute groupings cannot adapt to diverse product categories; and (2) Preference Conflicts in Unified Modeling, where single-output models struggle to resolve contradictory image preferences.

Methodology: The OSMF Framework

To address these limitations, the authors propose One Size, Many Fits (OSMF), a unified framework designed to align diverse group-wise click preferences. The framework consists of two primary stages:

1. Product-Aware Adaptive Grouping (PAAG)

PAAG replaces rigid, static user clustering with a dynamic approach that integrates user attributes with product characteristics.

User Preference Modeling: The system models user click behavior through three branches: user attributes ( $A$ ), advertising images ( $I$ ), and product titles ( $T$ ). It employs encoders to extract semantic features from images and text. Crucially, it utilizes cross-attention mechanisms to create product-aware user representations ( $e_{u|t}$ and $e_{u|c}$ ), allowing user features to attend to specific product semantics and visual characteristics.
Group Feature Aggregation: Users are clustered based on these product-aware embeddings using K-Means. To capture intra-group diversity rather than relying on a single centroid, PAAG constructs rich collective representations ( $G_{s,k}$ ) by sampling multiple representative points from the cluster distribution (including the centroid and peripheral points at various percentiles). This ensures the group feature encapsulates both core traits and subtle variations.

2. Preference-Conditioned Image Generation (PCIG)

PCIG leverages the groups derived from PAAG to generate tailored images via a Group-aware Multimodal Large Language Model (G-MLLM).

G-MLLM Architecture: The model integrates group information by prepending a group token ( $e_{G_{s,k}}$ $e_{G_{s, k}}$ ) to the standard input sequence. It undergoes specialized pre-training on four tasks:
1. Group Analysis: Describing typical user profiles in natural language.
2. Behavioral Prediction: Predicting future clickers based on existing ones.
3. Product Comprehension: Predicting product titles from images.
4. Prompt Generation: Creating image prompts tailored to specific users.
Group-wise Preference Alignment (Group-DPO): To align the model with real-world engagement, the authors introduce a Group-aware Reward Model (GRM). The GRM evaluates pairs of advertising images conditioned on specific group embeddings to predict relative CTR performance. Using these preference pairs, the G-MLLM is fine-tuned using Group Direct Preference Optimization (Group-DPO). This process maximizes the probability of generating prompts that yield higher CTR for specific target groups, effectively resolving cross-group preference conflicts without requiring separate models for each group.

Key Contributions

The paper outlines three primary contributions:

PAAG: A dynamic user grouping mechanism that learns product-aware preferences and derives rich collective features through adaptive clustering and feature sampling, overcoming the limitations of static grouping in large-scale scenarios.
PCIG: A preference-conditioned generation framework that utilizes a pre-trained G-MLLM and Group-DPO to enable a unified model to generate tailored images for diverse user groups while resolving preference conflicts.
GAIP Dataset: The introduction of the Grouped Advertising Image Preference (GAIP) dataset, the first large-scale public dataset for group-wise image preferences. It contains approximately 600,000 user groups derived from 40 million users, providing comprehensive group-wise preference annotations that significantly exceed existing datasets in scale and granularity.

Experimental Results

The authors evaluated OSMF through extensive offline and online experiments:

Preference Modeling (PAAG): Compared against baselines (CACS, WIYD, JAC), OSMF achieved the lowest NDCG@5 (0.3066) and highest AUROC (0.6372), indicating superior ability to distinguish distinct group preferences and model user clicks.
Online CTR Performance: In a one-week online test on 10,000 high-exposure products, OSMF achieved a 5.479% relative improvement in CTR over a pretrained G-MLLM baseline and outperformed the state-of-the-art CAIG method (which improved by 3.425%). This demonstrates that explicitly modeling group-level preferences yields better commercial outcomes than aggregate optimization.
Reward Modeling: The Group-aware Reward Model (GRM) significantly outperformed general-purpose MLLMs and existing CTR reward models in Pair Accuracy (improving by 4.7% over CAIG), validating the importance of group context in preference estimation.
Qualitative Analysis: Visual results showed that OSMF successfully generates stylistically diverse images (e.g., vibrant vs. minimalist) tailored to specific user groups for the same product, confirming its ability to adapt to heterogeneous demands.

Significance

The paper claims that OSMF represents a significant shift from the "one-size-fits-all" paradigm in advertising image generation. By introducing a framework that dynamically aligns diverse group-wise preferences, it addresses the commercial inefficiency of generic ad generation. The authors emphasize that even marginal CTR improvements translate to substantial revenue gains in large-scale e-commerce. Furthermore, the release of the GAIP dataset is positioned as a foundational resource for future research into group-aware image generation, filling a critical gap in the availability of large-scale, real-world group preference data. The work demonstrates that balancing scalability with personalization through group-wise optimization is a viable and effective path for industrial-scale advertising.