Computationally efficient searching, browsing and retrieval of one or more objects in a video sequence are accomplished using learned generative models. The generative model is trained on an automatically or manually selected query sequence from a sequence of image frames. The resulting generative model is then used in searching, browsing or retrieval of one or more similar or dissimilar image frames or sequences within the image sequence by determining the likelihood of each frame under the learned generative model. Further, this method allows for automatic separation and balancing of various causes of variability while analyzing the image sequence. The generative models are based on appearances of multiple, possibly occluding objects in an image sequence. Further, the search strategies used include clustering and intelligent fast forward through the image sequence. Additionally, in one embodiment, a fast forward speed is relative to the current frame likelihood under the learned generative model.