Submitted to arXiv on July 28, 2026 (paper 2607.25537), “Visual prompt engineering for video models” comes from a Google DeepMind team including Robert Geirhos, Yuxuan Li, Thaddaeus Wiedemer, Neha Kalibhat, Zi Wang, Mani Malek, Oyvind Tafjord, Kevin Swersky, Been Kim and Priyank Jaini. The premise is a direct analogy. Text prompt engineering became essential because a language model is only as good as its prompt; video models are now becoming foundation models for visual tasks, so the authors ask whether the same leverage exists on the image side.
Their method, which they abbreviate VIPE, automatically modifies the task image before handing it to the video model. The worked example is a visual physics question - where does a ball land after passing through a set of obstacles - where an abstract, sketch-like scene is converted into a photorealistic version with a single call to an image editing model. Nothing about the model or the text instruction changes; only the pixels the model is asked to reason over.
The finding is that this improves video reasoning performance across tasks, and that for video models visual prompt engineering can be more effective than either classic text-based prompt engineering or test-time scaling. That last comparison is the one worth noting: rewriting the picture beat spending more inference compute on the same picture. The authors position VIPE as a simple, compute-efficient way to elicit better visual reasoning, and host example videos on a project page.
For a technical leader, this is a cheap-lever result. Teams building on multimodal models have largely imported the language-model playbook - better instructions, more reasoning tokens - and left the visual input untouched. If a preprocessing step that costs one image edit outperforms extra inference budget, it belongs in the standard pipeline for any vision-heavy application, from document understanding to inspection and robotics.