Shifting the Breaking Point of Flow Matching for Multi-Instance Editing

Flow matching models have recently emerged as an efficient alternative to diffusion, especially for text-guided image generation and editing, offering faster inference through continuous-time dynamics. However, existing flow-based editors predominantly support global or single-instruction edits and struggle with multi-instance scenarios, where multiple parts of a reference input must be edited independently without semantic interference. We identify this limitation as a consequence of globally conditioned velocity fields and joint attention mechanisms, which entangle concurrent edits. To address this issue, we introduce Instance-Disentangled Attention, a mechanism that partitions joint attention operations, enforcing binding between instance-specific textual instructions and spatial regions during velocity field estimation. We evaluate our approach on both natural image editing and a newly introduced benchmark of text-dense infographics with region-level editing instructions. Experimental results demonstrate that our approach promotes edit disentanglement and locality while preserving global output coherence, enabling single-pass, instance-level editing.

Modern AI systems can edit images based on text instructions (for example, changing the color of a chair or the style of a background). However, most current tools struggle when you want to make several independent changes at once: asking the system to "make the left chair red" and "make the right chair blue" simultaneously often causes the edits to interfere with one another, producing unintended results in the wrong parts of the image. We investigate why this happens: these AI models tend to mix all instructions together internally when deciding how to modify the image, so individual edits can "bleed" into regions they were never meant to touch. To fix this, we propose a new mechanism that keeps each text instruction tightly bound to its intended region of the image, preventing cross-contamination between simultaneous edits. The result is a system that can apply multiple independent text-guided edits in a single step, each staying neatly within its target area while the overall image remains coherent. We test our approach on natural images and on a new benchmark we introduce, consisting of information-rich graphics such as posters and infographics, where precise, region-specific edits are especially important. Our results show that the method clearly improves the independence and accuracy of multi-target edits compared to existing approaches.