InnoAds-Composer: Efficient Condition Composition for E-Commerce Poster Generation
InnoAds-Composer employs single-stage tri-conditional diffusion with importance routing, achieving 54.39 FID and 0.857 sentence accuracy, greatly improving e-commerce poster generation.
Key Findings
Methodology
This paper introduces InnoAds-Composer built on MM-DiT, utilizing a unified multi-condition tokenization that maps style, subject, and glyph conditions into a shared embedding space. It employs importance analysis to route each condition to the most responsive layers and timesteps, reducing sequence length and computational load. The Text Feature Enhancement Module (TFEM) fuses glyph images and OCR features to improve Chinese text rendering. Layer and timestep importance guides selective condition injection, combined with a decoupled attention mechanism that isolates condition self-attention from the main noisy latent path. A high-quality 80K dataset with multi-condition annotations supports end-to-end training and evaluation, demonstrating superior control and efficiency.
Key Results
- In experiments, InnoAds-Composer outperformed baselines with a CSD of 0.729, DINO score of 0.923, IoU of 0.972, FID of 54.39, and IR-Score of 1.036. The importance routing reduced sequence length by ~30%, lowering inference costs while maintaining high fidelity. Ablation confirmed that injecting conditions only at most responsive layers preserves control quality and enhances efficiency.
- Compared to models like Flux, PosterMaker, and Qwen-Image-Edit, our approach achieved the best scores across all metrics, especially in text accuracy and style consistency. The model demonstrated robustness across diverse styles and complex scripts, generating visually appealing, semantically coherent posters that meet industry standards.
- The results validate that dynamic importance-based routing and feature fusion significantly improve multi-condition control, enabling faster inference without sacrificing quality, suitable for real-world deployment in automated advertising workflows.
Significance
This work advances the field of conditional image synthesis by integrating multi-modal controls into a single diffusion framework, addressing the long-standing challenge of balancing control precision and computational efficiency. It provides a scalable solution for automated, high-quality e-commerce poster generation, reducing reliance on manual design and enabling rapid customization. The proposed methods open avenues for broader applications in multi-objective content creation, setting new benchmarks for controllability and efficiency in generative models, and fostering industry adoption of intelligent advertising tools.
Technical Contribution
The key technical innovations include a unified multi-condition tokenization strategy, importance-aware dynamic routing of conditions, and a decoupled attention mechanism that isolates static condition features from the noisy latent path. These enable efficient multi-objective control without exponential growth in sequence length. The construction of a large-scale, multi-condition annotated dataset further supports robust training and benchmarking. The combination of these techniques results in a diffusion-based generator capable of high-fidelity, controllable poster synthesis with significantly reduced inference costs.
Novelty
This is the first work to unify style, subject, and glyph controls within a single diffusion model via importance-based routing, effectively reducing computational complexity. Unlike prior multi-stage or multi-model approaches, it achieves end-to-end, fine-grained control in a single pass. The dynamic routing based on importance analysis ensures that only the most influential conditions are injected at each layer and timestep, a novel strategy that balances control fidelity and efficiency, filling a critical gap in multi-objective generative modeling.
Limitations
- While effective, the model's performance diminishes with highly diverse or conflicting style and content inputs, especially in multilingual or highly stylized scenarios, due to limited training data diversity. High-resolution generation remains computationally intensive, limiting real-time applications. The importance routing relies on static importance maps, which may not adapt well to dynamic scene changes, potentially affecting control accuracy in some cases. Future work should focus on expanding multilingual support, optimizing high-res synthesis, and developing adaptive importance mechanisms.
Future Work
Future directions include integrating user preferences for personalized generation, exploring adaptive importance routing that dynamically updates during inference, and expanding the dataset to cover more languages and styles. Additionally, combining this framework with reinforcement learning could further refine control fidelity and efficiency. The ultimate goal is to develop a fully autonomous, real-time advertising content creation system that seamlessly adapts to diverse industry needs and user preferences.
AI Executive Summary
In the rapidly evolving landscape of e-commerce, visual advertising plays a crucial role in attracting consumers. Traditional poster creation relies heavily on manual design, which is time-consuming and often inconsistent. Recent advances in diffusion models have opened new avenues for automated image synthesis, yet existing methods struggle to simultaneously control multiple aspects such as style, text, and subject fidelity within a single, efficient framework.
Addressing these challenges, InnoAds-Composer introduces a novel single-stage diffusion framework that enables precise tri-conditional control over background style, product subject, and textual content. Built upon the MM-DiT backbone, it employs a unified tokenization strategy that maps heterogeneous conditions into a shared embedding space, facilitating joint inference. A key innovation is the importance analysis that guides dynamic routing, ensuring each condition influences only the most responsive layers and timesteps, significantly reducing sequence length and computational complexity.
The model also features a Text Feature Enhancement Module (TFEM), which fuses glyph images and OCR-extracted features to improve Chinese text rendering accuracy—a critical factor for effective advertising. Extensive experiments on a newly constructed 80K dataset demonstrate that InnoAds-Composer surpasses existing methods across multiple metrics, including style consistency (CSD 0.729), text accuracy (Sen.Acc 0.857), and overall image quality (FID 54.39). The importance routing reduces inference costs by approximately 30%, making the approach suitable for industrial deployment.
This work marks a significant step forward in automated advertising content generation, combining high fidelity, controllability, and efficiency. Its ability to generate diverse, style-coherent posters with accurate textual information opens new possibilities for rapid, scalable marketing solutions. Despite current limitations in multilingual support and high-resolution synthesis, ongoing research aims to extend these capabilities, promising a future where intelligent, fully automated ad design becomes a standard industry tool.
Deep Dive
Abstract
E-commerce product poster generation aims to automatically synthesize a single image that effectively conveys product information by presenting a subject, text, and a designed style. Recent diffusion models with fine-grained and efficient controllability have advanced product poster synthesis, yet they typically rely on multi-stage pipelines, and simultaneous control over subject, text, and style remains underexplored. Such naive multi-stage pipelines also show three issues: poor subject fidelity, inaccurate text, and inconsistent style. To address these issues, we propose InnoAds-Composer, a single-stage framework that enables efficient tri-conditional control tokens over subject, glyph, and style. To alleviate the quadratic overhead introduced by naive tri-conditional token concatenation, we perform importance analysis over layers and timesteps and route each condition only to the most responsive positions, thereby shortening the active token sequence. Besides, to improve the accuracy of Chinese text rendering, we design a Text Feature Enhancement Module (TFEM) that integrates features from both glyph images and glyph crops. To support training and evaluation, we also construct a high-quality e-commerce product poster dataset and benchmark, which is the first dataset that jointly contains subject, text, and style conditions. Extensive experiments demonstrate that InnoAds-Composer significantly outperforms existing product poster methods without obviously increasing inference latency.