[SIGIR 2025] NextAds: Moving Beyond Retrieval to Servicing-Time Video Generative Advertising

TL;DR

The era of picking the "least-bad" ad from a pre-made folder is over. NextAds proposes a paradigm shift where video advertisements are synthesized in real-time, tailored precisely to a user's aesthetic taste and the context of what they are currently watching. By leveraging VLMs as "Directors" and Sora-like models as "Producers," it transforms advertising from a discrete selection problem into a continuous optimization task.

Background Positioning: This is a pioneering work that provides both a conceptual framework and empirical benchmarks for Generation-based Personalized Video Advertising (GPVA), moving the industry toward the elusive "high personalization + high nativeness" quadrant.

Problem & Motivation: The Static Inventory Trap

In current systems (think YouTube or TikTok), even the most "personalized" ad is just a video chosen from a small pool. This retrieval-based paradigm fails because:

1. Granularity: You can't pre-produce a video for every unique sub-culture or aesthetic.
2. Latency: If an ad sucks, you can't fix it until the next production cycle weeks later.
3. Intrusiveness: Standard "pre-roll" ads break the user's flow because they aren't visually or contextually matched to the host content.

The authors' insight is that Generative AI allows us to treat the Ad Creative as a fluid, editable entity that can be optimized in a continuous space during the millisecond-long bidding process.

Methodology: The Four Pillars of NextAds

NextAds decomposes the complex task of video creation into a modular "agency-like" pipeline:

1. The Architecture

• Director: Uses a Vision-Language Model (VLM) to perform user-modeling and storyboard generation. It converts raw product data and user history into a frame-by-frame "blueprint."
• Producer: A video generation model (like Sora or Wan2.1) that renders the pixels based on the Director’s blueprint.
• Verifier: A gatekeeper that checks for "Hallucinations" (e.g., did the AI put a Nike logo on an Adidas shoe?) and safety compliance.
• Reflector: The "brain" that learns from skips or clicks to tweak the storyboard for the next impression.

2. Multi-Source Visual Grounding

To prevent the AI from "going rogue," the authors use Visual Collages. By combining product images, user-preferred colors, and the "host" video's first frame into a single visual prompt, they force the generator to stay faithful to the brand while remaining native to the environment.

Task Instantiations: PCG vs. PCI

The paper formalizes two key tasks:

1. Personalized Creative Generation (PCG): Creating a standalone "In-feed" ad.
2. Personalized Creative Integration (PCI): The "Holy Grail" of advertising—seamlessly weaving a product into an existing video (e.g., the host is cooking, and the ad smoothly introduces a specific brand of olive oil they are "using").

Experiments & Results

The authors tested their framework using a "Closed-source" stack (GPT-4o + Sora) and an "Open-source" stack (Qwen-VL + Wan).

• Personalization Win: In the PCI task, personalization scores for "Content" skyrocketed from 2.24 (Generic) to 8.32 (NextAds).
• Integration Quality: The Verifier-led integration point selection achieved a score of 8.67, meaning transitions felt natural rather than jarring.
• The Diversity Curse: Interestingly, as the ads became more personalized and integrated, their "Diversity" score dropped. This suggests that when an AI is told exactly how to be "native," it often converges on similar, safe narrative structures.

Critical Analysis & Conclusion

Takeaways

NextAds proves that we are no longer limited by the "Manual Production Bottleneck." We can now generate 1,000 different versions of a Coca-Cola ad for 1,000 different viewers, ensuring the aesthetic (Sichuan cuisine vs. Solo living) matches the user's current vibe.

Limitations & Future Work

• Spurious Conditioning: The AI sometimes injects "noise" (e.g., putting elements of a rival brand because it associated the user with that brand).
• Serving Latency: Rendering a Sora-quality video in the time it takes to load a webpage is currently impossible. The authors suggest "Predictive Caching" and "ROI-driven generation" (only generate if the expected lift justifies the GPU cost).

Final Thought: NextAds isn't just about making ads; it's about Progressive Storytelling. It envisions a future where an ad campaign evolves over a user's session, telling a coherent story across multiple impressions rather than just shouting the same slogan repeatedly.