Using Visual Search to Probe Veo 3.1

The project aims to use classical visual search paradigms to interrogate underlying mechanisms of generative video models. Specifically we will be using three 2 among 5 search task conditions to assess whether human-like attentional capabilities can be observed. The first condition will contain a basic feature, disjunctive, search which is expected to result in the relatively best performance of the model, mimicking pop-out effects that are described among human participants. The subsequent conjunctive conditions moderate difficulty along manipulations of the colour and from dimension. The second colour-conjunctive condition will interrogate the sensitivity of the model to colour conjunctions, where we expect a decline in generative accuracy relative to the first condition. The third condition contained both colour and form conjunctions illustrating the expected highest level of difficulty. Asides from the success rate hierarchy by conditions, we furthermore expect corresponding set size interactions, such that the disjunctive condition will not show effects of set size, but both the conjunctive conditions will result in decreased performance with increases in number of distractors, mirroring patterns observed in humans. In a direct to human comparison we expect relatively worse performance and additional measures observed are rates of generative fidelity which we expect to increase with increasing difficulty, that is condition and set size, of the generative task.