AutoPP: Towards Automated Product Poster Generation and Optimization
AutoPP integrates unified design, element rendering, and IDPO-based CTR optimization, achieving state-of-the-art automatic product poster generation and online CTR improvement.
Key Findings
Methodology
AutoPP employs a unified multimodal large language model (MLLM) to generate background, text, and layout simultaneously, utilizing flow matching for training. The element rendering module encodes product images and text via VAE and T5, applying Decomposed Attention to reduce computational load. The optimizer systematically replaces poster elements and uses Isolated Direct Preference Optimization (IDPO) to attribute CTR gains to individual elements, enabling fine-grained, targeted improvements. The entire pipeline is supported by AutoPP1M, a dataset with one million high-quality posters and user feedback, facilitating end-to-end automation and superior performance.
Key Results
- AutoPP outperforms prior methods like P&R and PAID on offline metrics such as FID (60.71) and CLIP-T (29.75), with layout alignment (Ali) at 0.007. Online CTR reaches 2.34%, a significant increase over baseline. Ablation shows Decomposed Attention reduces GFLOPs by 18%, and element-aware optimization yields more precise CTR attribution. The model generalizes well across diverse product categories, producing visually appealing posters with high textual accuracy (>95%).
- In large-scale online experiments, AutoPP's automatic element replacement and fine-grained CTR feedback lead to faster, more effective optimization compared to manual tuning. The system demonstrates robustness in real-world scenarios, with consistent CTR improvements and efficient deployment. Cross-scenario tests confirm its adaptability, making it suitable for industrial-scale applications.
- Ablation studies confirm that the combination of unified design, Decomposed Attention, and IDPO contributes significantly to performance gains. The method achieves a balance between computational efficiency and generation quality, with a 18% reduction in GFLOPs and a 1% higher CTR compared to holistic approaches. The element-specific attribution enables targeted improvements, leading to better user engagement metrics.
Significance
This work advances automated content creation by integrating multimodal generation and element-level optimization, addressing longstanding challenges in scalable, explainable CTR-driven advertising. It bridges the gap between offline quality metrics and real-world user engagement, providing a comprehensive framework for intelligent ad design. The large-scale AutoPP1M dataset further accelerates research in multimodal content generation and online learning, fostering industry adoption and academic exploration. The approach paves the way for personalized, adaptive advertising strategies that can dynamically respond to user preferences, significantly enhancing digital marketing efficiency.
Technical Contribution
AutoPP introduces a novel end-to-end pipeline combining a unified multimodal design model, an efficient element rendering mechanism with Decomposed Attention, and a systematic CTR optimization strategy based on IDPO. The integration of element-aware CTR attribution allows for precise, interpretable improvements, setting a new standard for automated content optimization. The use of large-scale datasets and advanced attention mechanisms demonstrates a significant leap over existing multi-stage or rule-based methods, offering both theoretical innovation and practical scalability.
Novelty
This research is the first to unify automated product poster generation and element-specific CTR optimization within a single framework. It innovatively applies Decomposed Attention to reduce computational complexity while maintaining high fidelity. The element-wise CTR attribution via IDPO is a pioneering approach that enables targeted, interpretable improvements, surpassing prior holistic optimization methods. The comprehensive dataset AutoPP1M further supports this innovation, establishing new benchmarks for multimodal content generation and online performance tuning.
Limitations
- Despite its strengths, the model's performance may decline in highly specialized or culturally specific scenarios where training data lacks diversity. The reliance on large-scale annotated datasets poses scalability challenges, especially for low-resource languages or niche markets. The current system's real-time inference speed still needs optimization for deployment in latency-sensitive applications. Additionally, the element attribution, while fine-grained, may not fully capture complex interactions among poster components, requiring further refinement.
Future Work
Future directions include extending the framework to multi-language and multi-cultural contexts, integrating reinforcement learning for dynamic CTR optimization, and enhancing real-time inference efficiency. Exploring personalized poster generation based on user preferences and browsing history is another promising avenue. Further research will also focus on improving element attribution accuracy and reducing dependency on large annotated datasets, making the system more adaptable and scalable for diverse industrial scenarios.
AI Executive Summary
In the rapidly evolving landscape of digital marketing, product posters serve as vital tools for capturing consumer attention. Traditional manual design processes are labor-intensive and lack scalability, especially given the demand for rapid content deployment across diverse product categories. Recent advances in deep generative models have enabled partial automation, but these methods often rely on multi-stage pipelines or handcrafted rules, which limit consistency and efficiency.
AutoPP introduces a comprehensive, end-to-end framework that automates both the generation and online optimization of product posters. Its core components include a unified multimodal large language model (MLLM) that jointly generates background, textual content, and layout, and a sophisticated element rendering module that encodes product images and texts into a compact latent space using Variational Autoencoders (VAE) and T5. The rendering process employs Decomposed Attention, a mechanism designed to efficiently handle long token sequences by separating intra-element dependencies from cross-element interactions.
The most innovative aspect of AutoPP is its online optimizer, which systematically replaces individual poster elements—background, text, or layout—and evaluates their impact on user engagement through Click-Through Rate (CTR) feedback. By employing Isolated Direct Preference Optimization (IDPO), the system attributes CTR improvements to specific elements, enabling targeted, interpretable refinements. This fine-grained approach significantly outperforms traditional holistic optimization methods, as demonstrated by experimental results showing a CTR increase to 2.34%, surpassing baseline models.
Supported by the AutoPP1M dataset—comprising one million high-quality posters and user feedback—AutoPP achieves state-of-the-art performance in both offline metrics (FID, CLIP-T) and online CTR improvements. Its ability to generate diverse, visually appealing posters with high textual accuracy and layout rationality makes it highly suitable for industrial deployment. While promising, the framework faces challenges such as adaptation to niche markets and real-time inference efficiency, guiding future research directions. Overall, AutoPP represents a significant step toward fully automated, intelligent advertising content creation, with broad implications for digital marketing and AI-driven content generation.
Deep Dive
Abstract
Product posters blend striking visuals with informative text to highlight the product and capture customer attention. However, crafting appealing posters and manually optimizing them based on online performance is laborious and resource-consuming. To address this, we introduce AutoPP, an automated pipeline for product poster generation and optimization that eliminates the need for human intervention. Specifically, the generator, relying solely on basic product information, first uses a unified design module to integrate the three key elements of a poster (background, text, and layout) into a cohesive output. Then, an element rendering module encodes these elements into condition tokens, efficiently and controllably generating the product poster. Based on the generated poster, the optimizer enhances its Click-Through Rate (CTR) by leveraging online feedback. It systematically replaces elements to gather fine-grained CTR comparisons and utilizes Isolated Direct Preference Optimization (IDPO) to attribute CTR gains to isolated elements. Our work is supported by AutoPP1M, the largest dataset specifically designed for product poster generation and optimization, which contains one million high-quality posters and feedback collected from over one million users. Experiments demonstrate that AutoPP achieves state-of-the-art results in both offline and online settings. Our code and dataset are publicly available at: https://github.com/JD-GenX/AutoPP