NextAds: Towards Next-generation Personalized Video Advertising

The gist: NextAds proposes a generation-based paradigm for next-generation personalized video advertising, shifting from static retrieval to dynamic, generation-based optimization in a continuous creative space.

Paradigm Shift: From Retrieval to Generation

NextAds moves personalized video advertising from static retrieval over a discrete creative inventory to dynamic generation-based optimization in a continuous creative space (Page 2). This paradigm shift is motivated by the limitations of existing retrieval-based systems, which suffer from Static inventory constraints and Feedback latency (Page 2). The goal is to optimize video creatives in a continuous space at serving time using Generative AI (GenAI) (Page 1).

NextAds Paradigm Principles

The NextAds paradigm is a generation-based, closed-loop paradigm for video creative optimization, empowered by GenAI (Page 3). It operates on several key principles:

1. Continuous optimization: Video creatives are continually synthesized and optimized with GenAI, rather than selected from a finite, pre-produced video pool (Page 3).

2. Plan-as-interface: A structured creative plan (e.g., user profile, storyboard) serves as the first-class intermediate interface to improve controllability, editability, and accountability (Page 3).

3. Grounded generation: Generated creatives must be grounded in eligible assets and verified product facts to ensure faithfulness and reduce hallucinations (Page 3).

4. Pre-deployment verification: Creatives must be adhere to strict constraints, such as platform policy, safety, privacy boundaries, and fairness (Page 3).

5. Dynamic refinement: Verification results and online interaction signals can be converted into actionable updates during serving (Page 3).

System Architecture

NextAds is conceptualized as a modular system with four core components: Director, Producer, Verifier, and Reflector (Page 3). The architecture is illustrated in Figure 3 (Page 4).

1. Director: This module generates a personalized and context-aware creative plan by performing User modeling and User-product matching to yield a structured creative plan (Page 4). The planning procedure involves:

** User modeling: Based on user specific-data, such as demographics and behavioral history, Director infers users’ content consumption preferences (Page 4). The process yields a structured user profile and can be continuously updated as new behaviors are observed (Page 4).**

2. Producer: This module executes the detailed creative plan by realizing the storyboard script with the selected assets, producing candidate creatives (Page 4). The authors treat video generation models as off-the-shelf tools and do not attempt to dissect or improve the script-to-creative realization process (Page 4).

3. Verifier: Acting as a utility-and-compliance gatekeeper, Verifier performs a dual assessment: utility verification and compliance verification (Page 4). Utility verification assesses Video quality, Personalization, Product alignment, Context alignment (Page 4), while compliance verification checks for Safety, Privacy and personalization boundaries, Fairness and bias (Page 4).

4. Reflector: This is the feedback-to-refine engine, which transforms assessments into minimal revisions (Page 3). It receives two types of feedback: structured rationales with localized evidence from Verifier and online user interaction signals (Page 3). Reflector performs attribution to distinguish between Plan-level issues and Execution-level issues to produce revision directives for the Director or Producer (Page 4).

Representative Tasks and Benchmarks

To facilitate comparable research, two representative tasks are formulated:

1. Personalized Creative Generation (PCG): This task aims to produce a standalone personalized video ads that are consumed independently as content (Page 4). Inputs include User-specific data, such as demographics and behavioral history; target product, including key selling points, and eligible assets like product images and logo; and serving context (Page 4). The objectives are to maximize advertising utility along three dimensions: video quality, personalization, and product alignment, while passing compliance verification (Page 4).

2. Personalized Creative Integration (PCI): This task focuses on seamlessly embeding advertising messages into an existing host video in a personalized and native manner without degrading the viewing experiences (Page 4). Inputs include the same as PCG plus a host video to be integrated with (Page 4). The objective is to optimize utility while preserving the host content experience, emphasizing integration quality, requiring low intrusiveness to the host video (Page 4).

To support systematic evaluation, two lightweight benchmarks are constructed: PCG-Bench and PCI-Bench (Page 4).

** PCG-Bench is built on the Qilin recommendation dataset where each user is associated with a historical interaction sequence of multimodal notes with metadata (title, image, and text) (Page 6). It selects 50 representative users with distinctive and consistent preferences and pairs them with five most compatible products from the product library (Page 6).**

** PCI-Bench is built on the MicroLens dataset which provides user interaction histories and multimodal video information, where it selects 50 representative users with clear preferences and pairs each with five best-matched products from the product library (Page 6).**

Feasibility Study and Performance

Exploratory experiments investigate the feasibility of NextAds, focusing initially on validating Director and Producer (Page 5). The implementation uses both open-source (Qwen3-VL5 as Director, Wan2.26 as Producer) and closed-source models (GPT-4o [18] as Director, Sora2 [30] as Producer) for comparison against GenericAd (Page 5).

Empirical results show that NextAds significantly improves personalization while preserving comparable product alignment and video quality compared to GenericAds in both PCG and PCI tasks (Table 1, Page 7). For instance, on PCG-Bench, NextAds achieves a Personalization score of 7.12 against GenericAds' 1.77 (Page 7). The case study demonstrates that the model performs adaptive user-product matching by prioritizing user–product compatibility over naive top-interest matching (Page 5).

Key Challenges and Opportunities

The paper discusses several critical challenges and future research directions:

1. Failure Analysis: Two failure modes are observed: Spurious conditioning, where models introduce unnecessary visual cues that act as noise, and Forced personalization under low compatibility, where forcing a match yields incoherent creatives (Page 5).

2. Key Insights in Pipeline Design: The paper highlights two pivotal architectural principles: Decoupling planning from execution to preserve narrative coherence, and using Multi-source visual grounding for faithful generation, combining target product images, user-preferred visual elements, and the integration start frame as hard constraints (Page 5).

3. Potential Research Opportunities: Future work should focus on:

** Continually user modeling across sequential impressions by applying causal and debiasing methods to disentangle genuine preference from exposure effects (Page 9). Developing a hierarchical memory for managing multitimescale user preferences is also suggested (Page 9). The trade-off between personalization and diversity can be mitigated by introducing controlled stochasticity to inject novel narrative styles, visual aesthetics, or unexpected hooks to avoid fatigue (Page 9).**

** Utilizing user simulators for offline iterative tuning and pre-deployment utility verification is proposed to bridge the gap between offline evaluation and online revenue (Page 9).**

** Balancing generative fidelity with serving latency requires exploring predictive caching, which anticipates user interests to pregenerate creative candidates during off-peak periods or ROI-driven generation, where systems explicitly optimize for Return-on-Investment (Page 9).**

4. Synergy: The paradigm shifts targeting from a filter to a mechanism that identifies high-value user–product pairs before using GenAI to synthesize personalized creatives (Page 9). This suggests research into Augment targeting with generatability prediction and designing auctions that trade off generation cost and expected lift (Page 9).

The NextAds paradigm offers a foundation for scalable, adaptive, and compliant personalized video advertising by moving toward dynamic, generation-based optimization (Page 8).

1. Bold header: Continuous user modeling across sequential impressions

This improvement involves applying causal and debiasing methods to disentangle genuine preference from exposure effects to move beyond fixed profiles, ensuring longitudinal personalization by jointly capturing short-term intent and long-term interests.

2. Bold header: Exploration strategy for diversity mitigation

To counter the risk of over-optimization, the system should adopt controlled stochasticity within the generative process to inject novel narrative styles or visual aesthetics, thereby addressing both fatigue mitigation and continual user preference refinement.

3. Bold header: Pre-deployment utility verification integration

The Verifier module can be enhanced by incorporating a user simulator during the pre-deployment phase to perform pre-deployment utility verification, which allows for validation of candidate creatives before online deployment, addressing the gap between offline metrics and online revenue.

4. Bold header: ROI-driven generation mechanism

To balance fidelity and latency, systems should implement an ROI-driven generation strategy where the system explicitly optimize[s] for Return-on-Investment (ROI) when deciding whether to invoke GenAI, reserving expensive generative processes only for impressions where personalization yields the highest marginal return.

5. Bold header: Augmenting targeting with generatability prediction

Targeting mechanisms should be augmented by predicting whether personalization is feasible for a given user–product pair, i.e., the expected headroom of generation under available assets and contextual constraints, ensuring that ads are only targeted when high-quality personalized creatives can actually be synthesized.

6. Bold header: Multi-source visual grounding for faithful generation

The system should utilize three complementary visual references (target product images, user-preferred visual elements, and integration start frame) as hard constraints to improve product faithfulness, preference alignment, and integration coherence, mitigating the risk of spurious conditioning.