UniEdit-I: Training-free Image Editing for Unified VLM via Iterative Understanding, Editing and Verifying

TL;DR

This work tackles the misalignment between semantic understanding and pixel-level generation in unified vision-language models by introducing UniEdit-I, a training-free, closed-loop editing framework that operates entirely in the semantic latent space. It leverages an Understanding–Editing–Verifying loop to translate instructions into semantically faithful edits with real-time VLM feedback and adaptive editing gain, enabling self-correcting trajectories. Evaluations on GEdit-Bench show state-of-the-art performance without fine-tuning and reveal artifact-free intermediate states that support stable feedback. The approach highlights the importance of semantic latent spaces for reliable, controllable image editing and paves the way for reflectively guided generative systems.

Abstract

While Unified Vision-Language Models promise to synergistically combine the high-level semantic understanding of vision-language models with the generative fidelity of diffusion models, current editing methodologies remain fundamentally decoupled and open loop performing static, pre-defined transformations without dynamic feedback between semantic interpretation and visual generation. A central limitation stems from the representation gap: understanding typically leverages high-level, language aligned encoders, whereas generation relies on low level, pixel-space autoencoders, resulting in misaligned feature spaces. To bridge this gap, Recent advances such as Representation Autoencoders and BLIP3-o advocate performing diffusion-based modeling directly in high level features from pretrained semantic encoders. We find editing in the semantic latent space modifies conceptual representations rather than pixels, ensuring intermediates that are both semantically coherent and visually plausible. Building on this insight, We propose UniEdit-I, the first training-free, closed-loop image editing framework that operates entirely within the semantic latent space of a unified VLM by introducing an Understanding-Editing-Verifying (UEV) loop, By transforming the VLM from a posthoc evaluator into an in-process conductor, UniEdit-I establishes the first semantics-driven, self-correcting closed-loop image editing pipeline. Evaluated on GEdit-Bench, UniEdit-I achieves state of the art performance without any fine tuning or architectural modifications, and even surpasses several largescale pretrained editors.

Figures (23)

• Figure 1: The illustration of UniEdit-I. We introduce a novel training-free framework named UniEdit-I to enable the unified VLM with image editing capability via three iterative steps: understanding, editing, and verifying.
• Figure 2: (a) In pixel space, intermediate outputs exhibit a superposition of source and target content, resulting in visible ghosting and unnatural transitions.(b) In semantic space, intermediate states are clean and realistic, with coherent structure and no artifacts, leading to a natural and faithful final result.
• Figure 3: Structured Prompt Generation pipeline. Visual analysis→semantic decomposition→instruction mapping→target construction.
• Figure 4: Semantic trajectory editing in CLIP space(right to left). We build on FlowEdit’s $\Delta V_t$, but apply it adaptively using VLM feedback to avoid artifacts ($\times$) and stop at the correct target ($\checkmark$).
• Figure 5: UniEdit-I outperforms FlowEdit variants by adapting both intensity and duration. (a) Source image; (c–f) FlowEdit with fixed gain ($\alpha=1.0$) under different $[n_{\max}, n_{\min}]$ settings, all suffering from over- or under-editing; (b) Our method, using real-time feedback to stop early and preserve semantics. Only Ours achieves faithful, artifact-free editing without manual tuning.