Constrained Edit Fields for Training-Free Flow Editing

Jingxuan Kang; Yinsong Wang; Che Liu; Chen Qin

cs.CVSep 27, 2026

Constrained Edit Fields for Training-Free Flow Editing

Authors: Jingxuan Kang, Yinsong Wang, Che Liu, Chen Qin

Organizations: Imperial College London

Abstract

Text-guided image editing aims to perform a desired edit while preserving source content unrelated to it. Pretrained rectified-flow models enable training-free editing of real images through modifications to their sampling trajectories. However, responses at locations unrelated to the desired edit can still accumulate along the editing trajectory and become visible in the final result. To overcome this, we propose Constrained Edit Fields (CEF), which assigns each spatial location a continuous edit responsibility that quantifies its relevance to the desired edit. CEF estimates edit responsibility directly from the source image when the relevant content is present. For edits whose target content is absent from the source, CEF first generates an unconstrained proposal to reveal its realized spatial support and then estimates responsibility from that proposal. At each editing step, CEF decomposes the base edit field into prompt-induced and trajectory-induced components, enabling edit responsibility to preserve instruction-relevant updates while suppressing unintended trajectory-induced changes. Evaluated on all 700 PIE-Bench examples, CEF achieves state-of-the-art Structure Distance, background LPIPS, and background MSE with both Stable Diffusion 3.5 Medium and FLUX, while retaining competitive instruction alignment. On Stable Diffusion 3.5 Medium, it reduces these metrics over the previous best results by 10.2%, 21.2%, and 48.0%, respectively.

Figures & tables

Figure 1: Effect of constraining the edit field. The unconstrained edit field introduces updates across the latent grid, closing the subject’s eyes while substantially altering his appearance. CEF constrains these updates according to edit responsibility, completing the desired edit while preserving the source appearance.

Figure 2: Accumulation of unintended changes. (a) Cumulative field displacement outside the edited region, relative to the total unconstrained update inside it. CEF suppresses it while preserving most edited-region updates. (b) Pixel differences along the unconstrained trajectory reveal the intended edit and progressively emerging facial changes.

Figure 3: Overview of Constrained Edit Fields. Source or proposal evidence localizes the content involved in the edit and yields a continuous edit responsibility. At each step, the base edit field is exactly decomposed into prompt-induced and trajectory-induced changes. CEF uses the prompt-induced direction to form prompt-aligned and residual trajectory components, applying responsibility more strongly to the residual to construct the constrained edit field.

Figure 4: CLIP and PSNR on SD3.5. FlowEdit and CEF are evaluated with

n_{\max}\in\{15,22,33\}

Method	Model	Structure	Background Preservation				CLIP Similarity
Method	Model	Dist. $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MSE $\downarrow$	SSIM $\uparrow$	Whole $\uparrow$	Edited $\uparrow$
IP2P	SD1.5	58.13	20.95	159.20	230.87	76.39	23.61	21.77
P2P	SD1.5	15.44	27.52	59.69	34.20	84.41	24.75	21.01
RF-Inversion	FLUX	41.17	20.86	187.01	120.12	71.21	25.08	22.39
RFEdit	FLUX	25.15	24.33	121.59	56.98	82.84	25.57	22.54
FireFlow	FLUX	27.40	23.11	128.46	70.75	81.30	26.13	22.87

Table 1: Quantitative comparisons on PIE-Bench. Baseline aggregates except SAM-Flow are reproduced from Table 1 of DirectEdit ( Yang and Ye 2026 ) ; the SAM-Flow and CEF rows are evaluated by us. Structure distance and LPIPS are reported at

10^{3}

scale, MSE at

10^{4}

scale, and SSIM at

10^{2}

scale. Bold indicates the best result within each backbone.

Figure 5: Qualitative comparison on PIE-Bench. The same edits are compared under fixed configurations on SD3.5 Medium and FLUX. CEF preserves source appearance while realizing shape changes, attribute edits, object replacements, and object additions across both backbones. Red circles mark incomplete target edits or unintended changes in regions outside the edit.

Config.	Struct.	Background				CLIP
Config.	Dist. $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MSE $\downarrow$	SSIM $\uparrow$	Whole $\uparrow$	Edit $\uparrow$
w/o $r$	10.98	32.08	26.63	13.51	92.75	25.66	23.10
w/o prop.	12.49	31.66	32.22	18.11	92.09	25.76	23.14
CEF (Ours)	9.91	32.72	24.72	11.26	92.98	25.52	22.91

Figure 6: Component ablation on PIE-Bench with SD3.5 Medium. Each ablation independently removes one component from CEF.

Figure 7: Effect of replay. Both final constrained edits are run with the same realization $\Omega$ . With replay, responsibility $r_{\Omega}$ is estimated from the proposal generated with $\Omega$ , as shown by the red mask in the first column.

Figure 8: Effect of decomposition. The decomposed field gives lower Structure Distance, LPIPS, and MSE.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Runtime (s) $\downarrow$	Peak memory (GiB) $\downarrow$
FlowEdit	$3.5736\pm 0.0023$	26.13
CEF	$5.1050\pm 0.0040$	26.23

Appendix

Table A1: Editor runtime and peak allocated memory on SD3.5. Each timed run covers source and prompt encoding, the editing trajectory, and decoding, with a precomputed edit responsibility provided to CEF.

Backbone	$\boldsymbol{\alpha}_{\mathrm{s}}/\boldsymbol{\alpha}_{\mathrm{p}}$	$\boldsymbol{\gamma}_{\mathrm{s}}/\boldsymbol{\gamma}_{\mathrm{p}}$	Dist. $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MSE $\downarrow$	SSIM $\uparrow$	CLIP-W $\uparrow$	CLIP-E $\uparrow$
SD3.5 ( $n=50$ )	1.8 / 0.8	1.5 / 1.0	6.951	34.670	22.181	7.657	93.631	25.211	20.833
SD3.5 ( $n=50$ )	2.3 † / 1.0 †	1.5 † / 1.0 †	7.538	34.155	23.223	8.640	93.525	25.511	21.067
SD3.5 ( $n=50$ )	2.8 / 1.2	1.5 / 1.0	8.214	33.787	24.014	9.396	93.452	25.612	21.085
SD3.5 ( $n=50$ )	2.3 / 1.0	1.2 / 0.8	7.940	33.653	24.378	9.405	93.422	25.589	21.144
SD3.5 ( $n=50$ )	2.3 / 1.0	1.8 / 1.2	7.214	34.498	22.591	8.138	93.586	25.270	20.919
FLUX ( $n=50$ )	3.5	1.0	7.693	33.399	24.347	11.239	94.746	24.347	20.369

Appendix

Table A2: Sensitivity to

\alpha

and

\gamma

on the same 50-example subset used in Table 9 . Every row aggregates all 50 examples. Daggers mark the configurations used in the main experiments. Structure and LPIPS are reported at

10^{3}

scale, MSE at

10^{4}

scale, and SSIM at

10^{2}

scale.

Method	$\boldsymbol{n}_{\max}$	Dist. $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MSE $\downarrow$	SSIM $\uparrow$	CLIP-W $\uparrow$	CLIP-E $\uparrow$
FlowEdit	15	3.71	31.49	31.53	12.41	92.13	24.23	21.33
FlowEdit	22	7.60	28.59	46.14	22.59	90.37	25.33	22.22
FlowEdit	33	23.48	23.25	93.41	69.19	85.14	26.78	23.56
CEF	15	2.00	34.39	20.28	7.09	93.49	23.70	21.06
CEF	22	3.86	33.81	21.58	8.03	93.33	24.52	21.75
CEF	33	9.91	32.72	24.72	11.26	92.98	25.52	22.91

Appendix

Table A3: Results for FlowEdit and CEF on all 700 PIE-Bench examples at

n_{\max}\in\{15,22,33\}

. Structure and LPIPS are reported at

10^{3}

scale, MSE at

10^{4}

scale, and SSIM at

10^{2}

scale.

Figure A2: Two representative CEF failures on SD3.5. The ground-truth (GT) edit regions are included solely as post-hoc diagnostic references. Responsibility is shown in red; raw and constrained edit-field path lengths are shown in orange with a shared scale within each example.

Explore similar work

h-Flow: Flexible Flow-based Image Editing via Doob's h-Transform
Jul 12, 2026Zehui Guo, Zhen Wang, Junwei Shu +3Image Editing Text-To-Image Diffusion Models
East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China
DecFlowEdit: Self-Localized Flow-based Image Editing via Guidance Decoupling
Sep 28, 2026Zheyuan Zhan, Can Wang, Jiawei Chen +4Diffusion-Based Image Editing Image Editing
Zhejiang University · University at Buffalo · University of California, Berkeley
SAM-Flow: Source-Anchored Masked Flow for Training-Free Image Editing
Jun 4, 2026Haowang Cui, Rui Chen, Tao Luo +3Diffusion-Based Image Editing Image Editing
Tianjin Key Laboratory of Imaging and Sensing Microelectronic Technology, School of Microelectronics, Tianjin University, Tianjin 300072, China · School of Cyber Security, Tianjin University, Tianjin 300072, China

Jul 12, 2026, cs.CVJ/K move · Enter open · S save

h-Flow: Flexible Flow-based Image Editing via Doob's h-Transform

Zehui Guo, Zhen Wang, Junwei Shu, Yang Li+2

East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China

Image Editing Text-To-Image Diffusion Models Value Alignment

Open paper AskPDF arXiv

Editing images with pre-trained text-to-image flow models typically requires carefully balancing target alignment with the desired prompt and source consistency with the original image. Existing approaches either rely on inversion-based pipelines or heuristic source-to-target trajectory constructions, which often depend on architecture-specific designs or are sensitive to hyperparameters. In this paper, we propose h-Flow, a training-free and theoretically grounded flow-based editing framework. Inspired by Doob's

h

-Transform, we reformulate image editing as conditional generation under multiple terminal events corresponding to source consistency and target alignment. We first extend the classical

h

-Transform from SDE-based models to the deterministic RF framework by constructing an equivalent SDE with identical marginals. Within this formulation, we design dedicated

h

-functions for source consistency and target alignment, yielding closed-form reconstruction guidance and velocity-based semantic editing signals. We further introduce a velocity orthogonal decomposition to decouple reconstruction and editing directions, enabling a controllable trade-off between the two objectives. Extensive experiments demonstrate that h-Flow achieves effective, robust, and flexible editing across diverse scenarios. The code will be released soon.

Abstract

Explore similar work

h-Flow: Flexible Flow-based Image Editing via Doob's h-Transform

DecFlowEdit: Self-Localized Flow-based Image Editing via Guidance Decoupling

SAM-Flow: Source-Anchored Masked Flow for Training-Free Image Editing