DiTailed: Ensuring Visual Object Consistency in Text-Image-to-Image Flow Matching Models

2026-07-14

AI Summary

Text-guided image editing often struggles to preserve visual consistency due to the loss of key object attributes. To address this challenge, this work introduces ABO-Edit, the first high-quality dataset specifically designed for evaluating object-level consistency in image editing. The study further reveals that the conditional embedding space of flow-matching models retains implicit information about the target image even under high noise levels. Building on this insight, the authors propose FlowMirror, a parameter-free auxiliary loss that enhances object consistency during editing without requiring any architectural modifications. Experimental results demonstrate that FlowMirror significantly outperforms existing baselines across multiple metrics, achieving notable improvements in preserving both structural integrity and semantic attributes of the edited subject.

Abstract

Despite remarkable progress in text-guided image editing, generative models frequently fail to preserve visual object consistency, defined as the preservation of a subject's key attributes throughout the editing process. We address this limitation through three contributions. First, we introduce ABO-Edit, a dataset specifically designed to study object consistency, comprising over 12,000 triplets of source images, editing prompts, and high-quality target images rendered from artist-designed 3D assets, with multi-view coverage and human-verified quality control. Second, we uncover an overlooked property of image-editing rectified flow models: the conditioning embedding space, not directly supervised during training, encodes a prediction of the final generated image even at high noise levels. Third, exploiting this finding, we propose FlowMirror, a parameter-free auxiliary loss that supervises this conditioning embedding space. Without architectural changes, our method improves generation quality across several metrics over baselines.

Problem

• visual object consistency
• text-guided image editing
• generative models
• image editing
• object preservation

Innovation

• object consistency
• rectified flow models
• conditioning embedding space
• FlowMirror
• ABO-Edit dataset