Text-to-Image Models Need Less from Text Encoders Than You Think
Researchers tested three contextless text embeddings—Bag-of-Tokens (BoT), Bag-of-Words (BoW), and Bag-of-Position-Tagged Words (BoPTW)—to see if diffusion models need rich text context. Findings show that while simple prompts work with BoT, complex ones benefit most from BoPTW. This suggests image models can interpret linguistic structure directly, reducing reliance on text encoders for full context.