Intelligent Creative Generation and Optimization for Xiaohongshu Advertising

Xiaohongshu’s advertising ecosystem relies heavily on user‑generated content (UGC). To improve ad performance while respecting community aesthetics, the platform has built an end‑to‑end intelligent creative system that lowers the production threshold, generates large volumes of high‑quality ad creatives at low cost, and performs efficient creative selection.

The system first extracts high‑quality image assets from rich media notes (cover images from image notes or key frames from video notes) and then generates multiple candidate titles. Title generation uses two paradigms: a GPT‑2 based model that consumes the full note text and conditional control signals (keywords, length, emojis) to produce titles closely related to the original content, and a Seq2Seq (T5) model that rewrites the original title by injecting high‑frequency queries, brand information, and style prompts to increase diversity.

To address the domain gap between public pretrained models and Xiaohongshu’s unique style, a family of RED pre‑training models was constructed on 1 billion internal notes: RED‑BERT for understanding, RED‑GPT2 for generation, and RED‑T5 for seq2seq tasks. These models capture platform‑specific vocabularies, emojis, and stylistic nuances.

Creative selection is framed as a Multi‑Armed Bandit (MAB) problem. An Upper Confidence Bound (UCB) strategy is adopted, where each creative’s reward is a weighted combination of CTR uplift, revenue increase, and a QoE‑derived bonus based on average view time. The reward formula incorporates a dynamic coefficient to balance user experience and advertiser goals.

To reduce exploration cost, an exit mechanism is introduced: once a creative’s CTR stabilizes within a predefined range, only the highest‑reward creative continues to be shown, while a sliding‑window ensures minimum exposure for all candidates.

Because raw exposure data can be biased by ad slot position, a position‑corrected metric called ECOI (Expected Click on Impression) is proposed, extending the COEC concept to align CTR across slots without additional hyper‑parameters.

For long‑tail creatives lacking sufficient feedback, a large‑scale discrete‑value DNN model provides generalized predictions. A dual‑tower architecture further mitigates computational explosion: the left tower predicts overall ad CTR using ad‑level embeddings, while the right tower predicts creative‑level CTR using creative embeddings and the left‑tower’s hidden state. During inference, the left tower runs once per ad, and the right tower runs for each creative, achieving linear scalability.

Feature engineering includes SENet‑weighted embeddings for ID features, regularization and adaptive dropout for over‑fitting mitigation, and cross‑features that combine user interests (derived from historical click sequences, OCR tokens, and entity terms) with creative attributes. This enhances both generalization and personalization, yielding significant uplift in CTR and revenue while preserving user QoE.

After more than a year of iteration, the intelligent creative platform now supports image and title generation, multi‑style rewriting, UCB‑based selection, position‑corrected metrics, and scalable dual‑tower inference, forming a solid foundation for future multimodal understanding and further optimization.