Cross-Attention Image Editing Without Masks for Structure Preservation

The patent introduces cross-attention maps as an intermediary mechanism that automatically identifies and isolates semantic regions of interest based on text prompts. Instead of requiring users to manually create masks, the system uses the cross-attention maps generated during image generation to automatically segment and control specific regions, thus achieving precise editing without cumbersome masking operations

The patent replaces the mechanical masking process (manual user interaction to define regions) with a computational approach using cross-attention maps. The system substitutes the manual mask creation mechanism with an automated attention-based region identification system that derives semantic regions directly from the text prompt and generated image, eliminating the need for manual masking operations

The patent performs preliminary action by preserving the full original image and its structural information throughout the editing process. Instead of removing or discarding the original image structure when creating masks, the system maintains the complete image and selectively applies modifications only to identified regions using the cross-attention maps, thus preventing loss of structural information

The patent applies local quality by making targeted modifications only to specific semantic regions identified through cross-attention maps, while leaving the rest of the image structure intact. This allows regional control to be exercised precisely where needed without affecting or removing global structural information from the original image

The patent performs preliminary action by generating cross-attention maps during the initial image generation process. These maps capture the relationship between text tokens and image regions, establishing a stable foundation that can be reused for editing. This preliminary capture of semantic relationships ensures that subsequent edits based on text prompts will reliably affect the intended regions without causing unpredictable changes

The patent implements feedback by using the cross-attention maps to guide the editing process. The system continuously references the pre-computed attention maps to ensure that text prompt changes are applied to the correct semantic regions, providing feedback control that prevents drift and ensures reliable, consistent editing outcomes even when prompts are modified

The patent applies segmentation by dividing the image into distinct semantic regions based on the cross-attention maps, each corresponding to specific text tokens. This segmentation enables precise control over different regions independently, allowing the system to edit specific objects or areas with high precision while maintaining ease of operation through text prompts

The patent adds another dimension to the editing process by incorporating cross-attention map guidance alongside text prompts. This additional dimension of control (the attention dimension) enables the system to maintain both ease of operation through natural language and high precision through spatially-aware region control, overcoming the limitations of simple inpainting methods