Datasets:

taesiri
/

arxiv_audio

+ "In a joint vision-language space, a text feature (e.g., from 'a photo of a dog') could effectively represent its relevant image features (e.g., from dog photos). Inspired by this, we propose PromptStyler which simulates various distribution shifts in the joint space by synthesizing diverse styles via prompts without using any images to deal with source-free domain generalization. Our method learns to generate a variety of style features (from 'a style of a') via learnable style word vectors for pseudo-words. To ensure that learned styles do not distort content information, we force style-content features (from 'a style of a [class]') to be located nearby their corresponding content features (from '[class]') in the joint vision-language space. After learning style word vectors, we train a linear classifier using synthesized style-content features. PromptStyler achieves the state of the art on PACS, VLCS, OfficeHome and DomainNet, although it does not require any images and takes just approximately 30 minutes for training using a single GPU."