Results showing vanilla WildCAT3D inference for novel view synthesis with various hard-coded trajectories including lateral turns, zoom-outs, and NeRF-like circular paths. Each result shows the input view and video showing generated novel views.
Results demonstrating appearance control during generation. By injecting appearance embeddings from external images, WildCAT3D can generate novel views while editing scene appearance (e.g., generating nighttime scenes from daytime photos). Each result shows the appearance conditioning image, generated view, and input view for comparison.
Results showing interpolation between views of scenes with differing appearances. WildCAT3D fine-tuned with two observed input views can interpolate between them to produce static videos with consistent appearance. Two interpolations are provided for each scene: one using the start view's appearance and another using the end view's appearance throughout the trajectory.