We can quickly categorize scenes into different classes, an ability that has been previously related to usage of both object information and global scene properties. Here we compared the utility of both sources of information and assessed the time course of information usage by reducing real-world scene images to either single objects or visualizations of global scene information (textures). In Experiment 1, using a 16AFC scene categorization task, we showed that object information was on average more useful than global scene information, and that combining both types of information (i.e., a single object in full resolution with a scene texture background) was not sufficient to explain fast scene categorization. In Experiment 2, we created inconsistent object-texture combinations using an object from one scene category and a texture from another, thus forcing participants to choose between the scene category conveyed by object vs. global scene information in a 2AFC task. In line with Experiment 1, we found a small but significant bias to choose the scene category conveyed by the single object over the category conveyed by global scene information. Surprisingly, however, this tendency decreased with longer SOAs between stimulus and mask, implying that participants increasingly based their categorization on global scene information when longer processing times were granted. Both these findings are inconsistent with the proposal of a default global-to-local processing sequence of scene information and could be interpreted as suggesting more flexible information usage during fast scene categorization. We discuss several alternative explanations and control experiments for our findings.