Steering Generative Video Models Toward Product Fidelity in Advertising

Oct 20 2:45 PM HST :calendar:
Audience level: Lightning Talk

About This Talk

Generative image and video models are rapidly entering ad-creative workflows, turning briefs into assets at scale. Yet product fidelity—accurate geometry, textures, colors, logos, and on-pack text—remains non-negotiable for customer trust, brand/legal compliance, and credible measurement. In this talk, I’ll outline a few fidelity-first approaches for steering generative video models toward product-faithful ad creative. I’ll discuss product-aware generation conditioned on a reference product, a model-agnostic post-restoration module that repairs drift from reference image(s), and a supervisory agent that automatically clips or discards distorted shots and, when needed, revises prompts or switches base models.

Presenters

    Photo of Baris Gecer, Phd

    Baris Gecer, Phd

    Senior Applied Scientist, Amazon Ads

    Baris Gecer is a Senior Applied Scientist at Amazon Ads advancing production-grade generative video for advertising and product content. A founding member of the team that built Amazon Ads Video Generator, he develops diffusion- and transformer-based pipelines for scalable, high-fidelity creation, and contributed to Nova Reel. His prior work spans photorealistic 3D face reconstruction, neural avatars, and telepresence across Huawei, FaceSoft, and Meta Reality Labs. He holds a PhD in Computer Science from Imperial College London. His research appears in CVPR, ICCV, ECCV and TPAMI and has led to granted patents. Beyond product impact, he contributes to open-source and reviews for leading computer vision venues.