How AIGC Transforms Advertising Material Creation on Xiaohongshu

Background

Advertising on platforms relies on creative assets (materials) that drive ROI for advertisers, retain users, and preserve community tone. Traditional pipelines involve costly manual steps from idea to shooting, making material production a major expense.

Why AIGC Matters

Generative AI offers a way to break the high‑cost, low‑efficiency status quo by automating parts of the creative workflow while still meeting three stakeholder needs: advertisers demand performance, platforms need quality, and users expect non‑intrusive experiences.

Industry Practices

Material production accounts for the largest share of advertising cost, especially on community‑centric platforms.

Advertisers pursue deterministic ROI, pushing for predictable, high‑quality assets.

Case Studies

An agency moved expert involvement earlier in the pipeline: idea → script → AIGC generation → expert review → shooting → online delivery.

Keyword‑rich titles were generated to improve search ranking.

A platform released an offline AIGC tool combined with online optimization.

SPU‑to‑material generation on the platform.

Fuzzy material content was used to deter crawling and drive external traffic.

Multimodal feature concatenation for online models.

Technical Solution

Controllable Generation Framework

The workflow consists of two stages:

Training: Auto‑Labeling + SFT (Supervised Fine‑Tuning) to build a controllable generation model.

Inference: Input a note, output promotion target + title sequentially.

Base Model Pre‑training

Billions of Xiaohongshu notes were cleaned and mixed with generic corpora to continue pre‑training a base model. A 10‑billion‑parameter model was selected after evaluating hallucination and title attractiveness, balancing performance with inference cost.

Automatic Labeling

Three controllable dimensions were defined: note content, audience segment, and title style. Synthetic data generated by a general LLM provided labels for promotion target, audience tier, and style, creating <note, keyword, title> and <note, style, title> pairs for downstream training.

Reward Models (RM)

Two RMs were built to filter generated titles:

Usability RM : Detects incoherent sentences, entity mismatches, and hallucinations using manually curated positive/negative samples and data‑augmented negatives.

Attractiveness RM : Predicts title appeal via pair‑wise loss on human‑ranked title lists, using the domain‑pretrained encoder as the backbone.

End‑to‑End Joint SFT & Inference

A unified label combines promotion target, audience tier, and multiple titles. The domain‑pretrained model is fine‑tuned end‑to‑end; at inference time, a note yields the promotion target followed by several controlled titles.

Evaluation

Diversity Evaluation

Baseline: random sampling from original note‑title pairs. Experimental model generated 12 titles per note, matching baseline quantity. Metrics included coverage, novelty, and click‑through simulations.

Consistency Evaluation

Human GSB (Ground‑Truth Subjective) assessment on 200 notes measured alignment between generated titles and marketing goals. Results showed a significant lift in consistency scores compared with the baseline.

Results and Insights

The AIGC pipeline achieved higher material diversity and quality while reducing production cost. Controllable generation plus RM filtering ensured titles remained on‑brand and attractive. Future work will explore multimodal features, preference learning, model compression, and broader advertising scenarios.