TORA

Rare concepts appear infrequently in training data, causing models to fail at visualization. However, their meaning is already latent within the text embeddings. By scaling up text embedding variance, the semantic spacing between tokens widens, making rare meanings more clearly visible. PCA-based principal component spaces are separated for "Token Spacing + Residual Alignment" to achieve balanced adjustment. TORA is a method that naturally extracts hidden rare meanings in Diffusion Transformers through simple embedding adjustments.