AdSelect: Teaching Multimodal LLMs to Clip Ads Like Professionals

Advertisers routinely create multiple duration variants of ads to meet marketing budgets and viewer preferences, a labor-intensive and costly process. We formulate ad shot selection as a set-prediction problem, mapping shots from the focal video to those retained in the target edit. To enable this, we introduce AdSelect, a dataset of ~4.8K long–short ad pairs mined from ~4M YouTube videos, split into ~4K training and 800 benchmark pairs. We discuss heuristic baselines that redefine the metrics and set a standard floor for performance on the task. Benchmarking open, frontier, and training-free models exposes a common failure pattern. They over-select while often missing the shots editors kept, hence no zero-shot model clears the baselines on our primary metrics. We propose AdCraft, a fine-tuning method combining LoRA with explicit complement-shot supervision to select shots that compose a high-quality short edit from a source. AdCraft reaches 0.771 precision and 0.688 IoU with duration error near one second, surpassing every heuristic baseline, zero-shot model, and prior supervised method. AdSelect together with AdCraft open new avenues for automated ad editing across platforms.

Producing a short ad from a long one means selecting and rearranging shots from the source by hand — labor-intensive, editor-dependent, and a major bottleneck in advertising production. We cast this editing step as a content-aware set-selection problem: given the shots of the source, select the subset that composes the short edit.

This framing reuses professionally produced footage rather than generating new content, and isolates a narrow, learnable skill — deciding which shots survive — instead of end-to-end generation. We condition on each pair's matched duration (the achievable length recovered from the retained shots, not a nominal 15s budget) and evaluate with precision and IoU as primary metrics, since recall-driven metrics reward indiscriminate over-selection.