Learning to Decompose Visual Features with Latent Textual Prompts