Generate color palettes from text prompts. A pretrained CLIP model is used to extract the text features, which are passed to a transformer decoder for conditional, autoregressive color generation. Noisy teacher forcing is employed in training to enhance robustness. Noise can be injected to the generated colors at intermediate steps to enhance palette diversity.

Generate natural-looking text morphing animations. Two text strings are converted into point sets, and a linear assignment algorithm matches corresponding points by minimizing the total squared Euclidean distance. This optimal matching reduces excessive movement, yielding fluid, visually pleasing transitions between them.

Flow-guided semantic word transitions in embedding space. Frequent words in a corpus are collected and the contextualized BERT embeddings are normalized onto a hypersphere. Related word pairs are identified using WordNet, and a vector field is then learned from optimal transport-induced displacements between the embedding distributions via Riemannian Conditional Flow Matching (RCFM). The continuous vector field is then utilized to guide a discrete graph search in the embedding space. This produces interpretable word sequences between any two given words (e.g., code → character → text → prose → poetry).