Negation without Confusion: Decomposing Bias and Recomposing Style in Text Embeddings for Negation-aware Image Generation

Key points

• To overcome this limitation, we propose a simple training-free method of Decompose and Recompose text embedding (Dero), which constructs a semantic-preserving subspace and a bias subspace using auxiliary prompts and performs semantic disentanglement based on bidirectional SVD.1
• By decomposing entangled representations and selectively recomposing disentangled semantic components, Dero realigns text representations toward intended negation semantics while restoring residual stylistic features to preserve semantic fidelity and image quality.1
• As a model-agnostic inference-time framework operating solely on text representations, Dero can be seamlessly applied to diverse diffusion architectures, including both U-Net and DiT based models.1
• Extensive experiments on both object- and concept-level negation demonstrate consistent improvements in negation-aware image generation, achieving stronger text-image alignment, better semantic fidelity, and higher human preference scores.1

Sources (1)

• [1]Negation without Confusion: Decomposing Bias and Recomposing Style in Text Embeddings for Negation-aware Image Generation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 01:40 AM

  Negation-aware image generation remains challenging for text-to-image diffusion models, which often struggle with negation due to implicit linguistic biases that are deeply entangled within text representations.

  To overcome this limitation, we propose a simple training-free method of Decompose and Recompose text embedding (Dero), which constructs a semantic-preserving subspace and a bias subspace using auxiliary prompts and performs semantic disentanglement based on bidirectional SVD.

Extractive summary: sentences quoted from the sources.