Segmentation-free guidance in diffusion models
By decomposing text cues into conditionalized lexical units in the diffusion model and adjusting the cross-attention weights, the problems of insufficient local fidelity and semantic consistency in image generation by the diffusion model are solved, achieving higher quality image generation while maintaining computational efficiency.
Patent Information
- Application Number
- CN202480069709.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-11-16
- Filing Date
- 2024-10-25
- Publication Date
- 2026-05-29
AI Technical Summary
Existing diffusion models suffer from insufficient local fidelity and semantic consistency when generating images, resulting in irrelevant or contradictory visual features in different spatial regions, which reduces image quality and alignment with the input text concepts.
By decomposing text prompts into multiple conditional terms and adjusting cross-attention weights during the iteration of the diffusion model, the most relevant conditional terms are determined for each image patch, reducing the influence of the highest semantic attention weight, achieving segmentation-free guidance, and ensuring precise alignment between image patches and text prompts.
It improves the spatial consistency and semantic fidelity of generated images, enhances image quality, makes them better match the intent of text prompts, and remains computationally efficient, requiring no additional classifier model.