Training-Free Image Editing with Visual Context Integration and Concept Alignment

Rui Song; Guo-Hua Wang; Qing-Guo Chen; Weihua Luo; Tongda Xu; Zhening Liu; Yan Wang; Zehong Lin

cs.CVApr 6, 2026

Training-Free Image Editing with Visual Context Integration and Concept Alignment

Authors: Rui Song, Guo-Hua Wang, Qing-Guo Chen, Weihua Luo, Tongda Xu, Zhening Liu, Yan Wang, Zehong Lin,

Organizations: HKUST · Alibaba International Digital Commerce Group · Institute for AI Industry Research (AIR), Tsinghua University · School of Data Science, Lingnan University

Abstract

In image editing, it is essential to incorporate a context image to convey the user's precise requirements, such as subject appearance or image style. Existing training-based visual context-aware editing methods incur data collection effort and training cost. On the other hand, the training-free alternatives are typically established on diffusion inversion, which struggles with consistency and flexibility. In this work, we propose VicoEdit, a training-free and inversion-free method to inject the visual context into the pretrained text-prompted editing model. More specifically, VicoEdit directly transforms the source image into the target one based on the visual context, thereby eliminating the need for inversion that can lead to deviated trajectories. Moreover, we design a posterior sampling approach guided by concept alignment to enhance the editing consistency. Empirical results demonstrate that our training-free method achieves even better editing performance than the state-of-the-art training-based models.

Figures & tables

Figure 2: The left figure shows the pipeline of FlowEdit. The middle figure illustrates the latent vectors, velocity fields, and sampling trajectory of VicoEdit. The right figure shows the procedure of each sampling step of VicoEdit.

Figure 3: Visualization of $\boldsymbol{z}^{tar}_{t}$ at different timesteps. We visualize the latents from two different trajectories, where the timesteps for starting sampling (i.e., $t_{n_{\text{max}}}$ ) are $0.93$ and $0.98$ , respectively. The model is instructed to replace the bear with the sloth. The visualization verifies that global features are generated at early steps, and skipping the early stage fails to alter the subject appearance.

Figure 4: Editing results with or without concept alignment. Concept alignment preserves details in the source image.

Figure 5: Visualization of $\boldsymbol{z}_{t}$ and $\hat{\boldsymbol{z}}_{0}$ at different timesteps. Concept alignment guidance accurately predicts $\boldsymbol{z}_{0}$ even at early timesteps (e.g., when $t=0.9$ ).

Figure 6: Source image, context image, and editing results of different methods.

Method	#Params	LPIPS ( $\downarrow$ )	DINO ( $\uparrow$ )	CLIP-T ( $\uparrow$ )
Diptych Prompting	12B	0.429	0.734	0.378
Qwen-2511	20B	0.521	0.679	0.381
FLUX.2	32B	0.345	0.683	0.381
Seedream 5.0 Lite	-	0.361	0.813	0.389
Nano Banana 2	-	0.378	0.786	0.387
VicoEdit (Ovis)	3B	0.515	0.720	0.387

Table 1: Comparison on model size and editing performance. Ovis, FLUX, and Qwen are abbreviations of Ovis-U1, FLUX.1-Kontext, and Qwen-Image. The best results are marked in bold , while the second best results are underlined .

Method	FLUX.2	VicoEdit (FLUX)	Qwen- 2511	VicoEdit (Qwen)	VicoEdit (Ovis)
Time (s)	193	122	83	184	81
Memory (GB)	$>$ 80	55	59	70	30

Table 2: Sampling time and peak GPU memory usage of VicoEdit and baseline methods.

Figure 7: Source, context, and edited images produced by VicoEdit and its variants.

Method	LPIPS ( $\downarrow$ )	DINO ( $\uparrow$ )	CLIP-T ( $\uparrow$ )
VicoEdit	0.322	0.731	0.378
Inversion	0.704	0.707	0.371
w/o Cpt Align	0.454	0.743	0.381
Skip Early Stage	0.147	0.427	0.332
Sample One Noise	0.362	0.730	0.378

Table 3: Ablation study on inversion-free sampler, concept alignment, and sampling strategies.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Model	$N$	$n_{\text{max}}$	$K$	$c^{src}_{T}$	$c^{tar}_{T}$	$c^{tar}_{I}$	$\tau$	$\alpha_{t}$
FLUX.1-Kontext	50	47	3	1.5	5.5 / 6.0	-	0.25	0.5
Qwen-Image-Edit	50	45	2	1.5	7.5 / 8.0	2.5 / 3.0	0.25	1.0
Ovis-U1	50	47	3	1.5	6.5 / 7.0	2.5 / 3.0	0.25	1.0

Appendix

Table 4: Hyper-parameters of VicoEdit.

c^{tar}_{T}

and

c^{tar}_{I}

columns showcase the CFG scales for replacement and add tasks.

Metric	Diptych Pmt.	FLUX.2	Qwen-2511	Seedream 5.0 Lite	Nano Banana 2	VicoEdit (FLUX)	VicoEdit (Qwen)	VicoEdit (Ovis)
SSIM $(\uparrow)$	0.346	0.657	0.451	0.652	0.605	0.744	0.611	0.488
CLIP-I $(\uparrow)$	0.899	0.873	0.881	0.941	0.931	0.909	0.896	0.905

Appendix

Table 5: Comparisons on SSIM and CLIP similarity. Diptych Pmt. denotes Diptych Prompting. FLUX, Qwen, Ovis represent the FLUX.1-Kontext, Qwen-Image-Edit, and Ovis-U1 version of VicoEdit. The best results are marked in bold , while the second best ones are underlined .

Metric	VicoEdit	$K=2$	$K=1$	$\tau=0.5$	$\tau=0.1$	$\alpha_{t}=0.25$	$\alpha_{t}=1$
LPIPS ( $\downarrow$ )	0.322	0.338	0.362	0.289	0.401	0.425	0.358
DINO ( $\uparrow$ )	0.731	0.729	0.730	0.666	0.730	0.735	0.733
CLIP-T ( $\uparrow$ )	0.378	0.379	0.378	0.370	0.379	0.380	0.380

Appendix

Table 6: Performance of VicoEdit under different hyper-parameter settings.

Figure 8: Results produced by FLUX.1-Kontext. The left half of each image pair shows the source and context images, while the right half presents the edited image.

Figure 9: Results generated by Qwen-Image-Edit. The left half of each pair shows the source and context images, while the right half corresponds to the editing result.

Figure 10: Editing results of Ovis-U1. The left column of each pair exhibits the source and context images, while the right column shows the editing result.

Concept Alignment

See all 13 papers

VeloEdit: Training-Free Consistent and Continuous Instruction-Based Image Editing via Velocity Field Decomposition
Mar 11, 2026Zongqing Li, Zhihui Liu, Yujie Xie +3Training-Free Image Editing Image Editing
Xiamen University · Truesight
Training-Free Multi-Concept Image Editing
Feb 24, 2026Niki Foteinopoulou, Ignas Budvytis, Stephan LiwickiTraining-Free Image Editing Image Editing
Cambridge Research Laboratory, Toshiba Europe, Cambridge, UK · Independent Researcher, UK
From Competition to Coopetition: Coopetitive Training-Free Image Editing Based on Text Guidance
Apr 17, 2026Jinhao Shen, Haoqian Du, Xulu Zhang +2Text-Guided Image Editing Attention Control in Diffusion Models
Department of Computing, Hong Kong Polytechnic University, Hong Kong, China · Department of Computer Science, Sichuan University, Chengdu 610017, China

Mar 11, 2026J/K move · Enter open · S save

VeloEdit: Training-Free Consistent and Continuous Instruction-Based Image Editing via Velocity Field Decomposition

Zongqing Li, Zhihui Liu, Yujie Xie, Shansiyuan Wu+2

Xiamen University · Truesight

Training-Free Image Editing Image Editing Text-Guided Image Editing Flow Matching

Open paper AskPDF arXiv

Instruction-based image editing aims to modify source content according to textual instructions. However, existing methods built upon flow matching often struggle to maintain consistency in non-edited regions due to denoising-induced reconstruction errors that cause drift in preserved content. Moreover, they typically lack fine-grained control over edit strength. To address these limitations, we propose VeloEdit, a training-free method that enables highly consistent and continuously controllable editing. VeloEdit dynamically identifies editing regions by quantifying the discrepancy between the velocity fields responsible for preserving source content and those driving the desired edits. Based on this partition, we enforce consistency in preservation regions by substituting the editing velocity with the source-restoring velocity, while enabling continuous modulation of edit intensity in target regions via velocity interpolation. Unlike prior works that rely on complex attention manipulation or auxiliary trainable modules, VeloEdit operates directly on the velocity fields. Extensive experiments on Flux.1 Kontext and Qwen-Image-Edit demonstrate that VeloEdit improves visual consistency and editing continuity with negligible additional computational cost. Code is available at https://github.com/xmulzq/VeloEdit.

Abstract

Concept Alignment

Explore similar work

VeloEdit: Training-Free Consistent and Continuous Instruction-Based Image Editing via Velocity Field Decomposition

Training-Free Multi-Concept Image Editing

From Competition to Coopetition: Coopetitive Training-Free Image Editing Based on Text Guidance