Text-to-image diffusion models have a well-known blind spot: they struggle to depict what a prompt says should not be there. A new arXiv preprint, “Negation without Confusion: Decomposing Bias and Recomposing Style in Text Embeddings for Negation-aware Image Generation,” argues that the difficulty is not simply a matter of parsing words. Implicit linguistic biases are deeply entangled within the text representations the models rely on, so negation cues can be lost or overridden.
To address this, the work centers on decomposing those biases and recomposing style in text embeddings. By separating bias from the stylistic content of a prompt, the approach aims to give negation a clearer path through the generation process. The abstract does not include quantitative results, but the framing points to a representation-level fix rather than a change to the diffusion architecture itself.
As the only source, the preprint offers a focused diagnosis of the problem and a proposed direction, without comparison to alternative methods in the available text.