Inverse-and-Edit: Simple and Effective Framework for Fast Image Editing

Ilia Beletskii; Andrey Kuznetsov; Aibek Alanov

cs.CVJun 23, 2025

Inverse-and-Edit: Simple and Effective Framework for Fast Image Editing

Authors: Ilia Beletskii, Andrey Kuznetsov, Aibek Alanov

Organizations: HSE University, AIRI · AIRI, Sber, Innopolis

Abstract

Diffusion-based text-to-image models have recently been distilled into few-step samplers and fast generators, reducing inference cost while maintaining high image quality. However, text-guided editing with such accelerated backbones remains challenging: inversion in a small number of steps is unreliable, local errors accumulate along the trajectory, and existing fast-editing methods typically rely on backbone-specific inversion procedures, which limits their transferability. In this work, we propose a simple and universal framework for text-guided image editing with arbitrary accelerated diffusion and flow-matching models. Instead of optimizing local, per-timestep objectives, we directly minimize a global editing objective over the entire inversion-generation trajectory. The framework fine-tunes only 1.5-7 % of the base-model parameters through two lightweight LoRA adapters (forward and backward), trained on fewer than 10 K triplets from the publicly available HQ-Edit dataset. We validate our method on three architecturally distinct backbones: consistency-distilled SD1.5 LCM-Dreamshaper, adversarial-distilled SDXL-Turbo, and rectified-flow-distilled Z-Image-Turbo. On PIE-Bench, our method achieves the strongest editing alignment within each backbone family while ensuring a strong level of content preservation. Inference adds no extra steps beyond the standard inversion and editing passes. The project page with visual results is available at iliabeletskii.github.io/inverse-and-edit.

Figures & tables

Figure 1: Qualitative comparison on Z-Image-Turbo (flow-based) against RF-Edit, FlowEdit, FireFlow, and UniEdit-Flow (inference time below each column; one edit per row). Our method applies each edit while preserving structure, and is the fastest editor at 3.18 s.

Figure 2: Overview of our framework. Left (fine-tuning): the forward adapter $f_{\phi}$ inverts $x_{0}$ to a terminal latent $z_{T}$ ; the backward adapter $g_{\theta}$ denoises $z_{T}$ under $y_{\mathrm{src}}$ (reconstruction, scored by LPIPS) and under $y_{\mathrm{trg}}$ (edit, scored by directional CLIP). Both adapters are trained jointly end-to-end with a prior regularizer $\mathrm{KL}+\mathrm{LF}$ on $z_{T}$ and the backbone frozen (Eq. 21 ). Right (inference): the same $f_{\phi}$ inverts, then guided $\hat{g}_{\theta}(z_{T},y_{\mathrm{trg}},\omega)$ produces the edit with no auxiliary machinery (Eq. 24 ).

Method	CLIP ${}_{\text{wh}}$ $\uparrow$	dCLIP $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	IR $\uparrow$
w/o $\mathcal{L}_{\text{KL}}$	$26.06$	$0.1350$	0.1901	0.8306	$21.81$	$0.6029$
w/o $\mathcal{L}_{\text{LF}}$	$26.13$	$0.1423$	$0.2071$	$0.8171$	$21.49$	$0.6015$
Ours (full)	26.44	0.1541	$0.1959$	$0.8296$	22.01	0.7383
w/o $\mathcal{L}_{\text{dCLIP}}$ (no edit signal)	$23.87$	$0.0457$	$0.0761$	$0.9075$	$25.72$	$-0.1320$

Table 1: Ablation of objective terms on Z-Image-Turbo (step 600); bold marks the best editing run . Without

\mathcal{L}_{\text{dCLIP}}

the model merely reconstructs the source (dCLIP

0.046

, IR

<0

); the full objective is best on editability and quality. LPIPS (VGG), SSIM, and PSNR are whole-image.

Figure 3: Qualitative comparison on SDXL-Turbo against ReNoise, GNRi, and TurboEdit (one edit per row). The baselines either lose input identity and detail or fail to apply the edit; our method applies each edit while preserving composition, identity, and background.

Figure 4: Qualitative comparison on LCM-DreamShaper against iCD, NPI, NTI, SwiftEdit, InfEdit, ChordEdit, and G&R (one edit per row). Most baselines barely edit the input or distort the scene when they do; our method produces a consistent edit per row while preserving layout, identity, and background.

	Editability			Preservation				Quality	Efficiency	Summary
Method	CLIP wh $\uparrow$	CLIP ed $\uparrow$	dCLIP $\uparrow$	Struct $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	IR $\uparrow$	Time $\downarrow$ (s)	Avg Rank $\downarrow$
SD-based models
G&R	26.47	22.79	$0.106$	13.7	$0.070$	$0.834$	$25.0$	$0.427$	$19.7$	3.33
InfEdit	$25.57$	$22.24$	$0.096$	$18.5$	0.053	0.850	28.0	$0.278$	$3.5$	3.33
ChordEdit	$25.90$	22.54	0.119	$29.7$	$0.122$	$0.764$	$22.4$	0.516	$3.5$	$5.11$
SwiftEdit	$24.61$	$21.34$	$0.081$	$16.2$	$0.077$	$0.820$	$24.0$	$-0.104$	0.2	$5.78$

Table 2: Baselines vs Ours under the official PIE-Bench protocol (CLIP wh , background-masked CLIP ed /LPIPS/SSIM/PSNR and Structure Distance

\times 10^{3}

); dCLIP is the directional CLIP score; ImageReward and time. Within each block, best is bold and second-best is underlined . Avg Rank is the mean dense-rank over the nine metrics, lower is better, with ties resolved on displayed precision.

Ours vs.	Overall $\uparrow$	Match $\uparrow$	Preserve $\uparrow$
RF-Edit	78.0	59.0	81.0
FlowEdit	76.0	58.0	81.0
FireFlow	81.0	58.0	83.0
UniEdit-Flow	70.0	59.0	70.0
Aggregate	76.3	58.5	78.8

Table 3: Human preference study on Z-Image-Turbo (flow-based). Pairwise comparisons of our method against each flow editor,

50

per baseline (

200

total). Each cell reports the preference score

\mathrm{win}+\tfrac{1}{2}\mathrm{tie}

(%);

50\%

is parity, higher favors ours. Overall is overall preference, Match is agreement with the edit prompt, and Preserve is fidelity to the source image.

SD-based methods				SDXL-based methods				Flow-based methods
Model	MSE $\downarrow$	ImageReward $\uparrow$	LPIPS $\downarrow$	Model	MSE $\downarrow$	ImageReward $\uparrow$	LPIPS $\downarrow$	Model	MSE $\downarrow$	ImageReward $\uparrow$	LPIPS $\downarrow$
DDIM-50	0.0079	-0.023	0.279	GNRi	0.0065	0.416	0.283	UniEdit-Flow	0.0127	-0.050	0.274
iCD	0.0447	-0.497	0.611	TurboEdit	0.0470	0.427	0.483	FireFlow	0.0092	-0.102	0.327
SwiftEdit	0.0496	0.437	0.507	ReNoise	0.0615	0.367	0.541	RF-Edit	0.0119	-0.342	0.376
Ours	0.0056	0.425	0.172	Ours	0.0061	0.504	0.186	Ours	0.0052	0.535	0.166

Table 4: Evaluation metrics for image reconstruction across SD-, SDXL-, and flow-based methods. Best ( bold ) and second best ( underlined ) are computed within each block for each metric. Only methods that perform an explicit inversion are included.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: The KL blind spot. A per-channel standardized image latent (middle; the bicycle is faintly visible) passes the channel-statistics test perfectly (KL $=0.000$ ) yet carries $16.9\times$ the white-noise low-frequency energy; our inverted $z_{T}$ (right) passes both tests.

Figure 6: Gradient-flow ablation: same recipe with gradients detached at

z_{T}

(Z-Image Turbo,

\omega{=}2.5

). Left: without through-cycle gradients, preservation never improves. Right: the truncated model’s higher dCLIP reflects weakly-anchored regeneration from the caption, not editing of the input. LPIPS-pres: VGG-16 LPIPS outside the edit mask.

Figure 7: Cycle length. Left: reconstruction LPIPS (VGG-16) vs. cycle steps

k

(PIE-Bench, n=50); errors do not accumulate, the trained model improves with

k

while the base backbone stays flat. Right three panels: retraining the same recipe at

T\in\{2,4,6,8\}

and evaluating each at its native

T

; editing saturates at

T{=}6

, preservation peaks at

T{=}8

(our setting, circled).

Figure 8: Causal LF-contamination experiment (PIE-Bench,

n{=}100

\omega{=}2.5

). Low-frequency noise is injected into

z_{T}

at increasing strength and renormalized per channel, so channel-KL is unchanged by construction; editing quality nevertheless degrades monotonically in dCLIP, preservation LPIPS, and ImageReward. The statistic penalized by the LF hinge is causally harmful, not merely correlated. Channel-KL here is the mean per-image statistic over the 100 PIE-Bench source images (right-skewed; median

0.006

) and is unchanged by the intervention by construction; the MS-COCO medians quoted in Sec. C.2 are computed identically on a different image set.

Panel of Figure 9 — Figure 9: The learned terminal latents compose and interpolate. Left: splicing the right half of $z_{B}$ into $z_{A}$ decodes each half near round-trip quality. Right: slerp interpolation between two inverted latents (decoded with an empty prompt) yields smooth semantic morphs.

Backbone	LoRA rank	Per adapter	Total trainable	Frozen base	% of base
SD1.5 (LCM)	32	6.38 M	12.75 M	$\sim$ 865 M	$\sim$ 1.5%
SDXL-Turbo	64	92.90 M	185.79 M	$\sim$ 2.6 B	$\sim$ 7%
Z-Image-Turbo	64	66.85 M	133.69 M	$\sim$ 6.2B	$\sim$ 2.17%

Appendix

Table 5: Trainable parameters of our two adapters, counted directly from the checkpoints. The forward and backward LoRA are identical in size; “% of base” is the total trainable fraction of the frozen backbone.

	CLIP wh $\uparrow$	CLIP ed $\uparrow$	dCLIP $\uparrow$	Struct $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	IR $\uparrow$	Time $\downarrow$ (s)
SD-based
PreciseInv	26.65	22.95	0.129	33.0	0.132	0.765	21.9	0.520	3.4
TFinv	24.53	21.48	0.072	45.0	0.133	0.744	19.8	0.036	165
Ours	26.13	22.47	0.155	22.8	0.071	0.843	23.6	0.573	0.6
SDXL-based
PreciseInv	27.49	24.08	0.151	64.6	0.213	0.735	18.2	0.736	9.4

Appendix

Table 6: Optimization-based few-step inversion vs. ours under the official PIE-Bench protocol (dCLIP and Image Reward from our scorer). Both baselines optimize per image at test time; ours is amortized. Time is per-image inference on the same H100 GPUs as Table 2 ; the baselines’ per-image optimization early-stops, hence faster than the nominal step count.

	CLIP wh $\uparrow$	CLIP ed $\uparrow$	dCLIP $\uparrow$	Struct $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	IR $\uparrow$
Ours	26.13	22.47	0.155	22.8	0.071	0.843	23.6	0.573
Ours (masked)	25.43	21.85	0.120	14.3	0.051	0.864	25.8	0.440

Appendix

Table 7: Masked-inference variant on the SD backbone (same checkpoint and

\omega{=}2.0

as Table 2 ); official PIE-Bench protocol, dCLIP and Image Reward from our scorer. dCLIP (

0.120

) still matches or exceeds every SD baseline (the closest, ChordEdit, scores

0.119

)

	SD		SDXL		Flow
Category	Ours	Best comp.	Ours	Best comp.	Ours	Best comp.
Random	2.33	2.44 (InfEdit)	1.89	2.11 (TurboEdit)	1.33	2.67 (UniEdit)
Change object	2.22	2.33 (InfEdit)	1.67	2.22 (TurboEdit)	2.33	2.67 (FlowEdit)
Add object	1.44	2.89 (InfEdit)	1.56	2.33 (TurboEdit)	1.56	2.67 (UniEdit)
Delete object	2.00	2.67 (InfEdit)	1.78	2.22 (GNRi)	$=$ 2.78	2.78 (FlowEdit)
Content	1.44	2.67 (InfEdit)	1.78	2.11 (TurboEdit)	1.56	2.67 (UniEdit)

Appendix

Table 8: Per-category Avg Rank vs. accelerated editors (lower is better): ours against the best competitor in each cell. Bold marks the cell winner;

=

marks ties.

SD-based			SDXL-based			Flow-based
Method	L/14	bigG	Method	L/14	bigG	Method	L/14	bigG
Ours	0.155	0.183	Ours	0.178	0.231	Ours	0.154	0.193
ChordEdit	0.119	0.145	TurboEdit	0.145	0.187	UniEdit-Flow	0.123	0.156
G&R	0.106	0.123	ReNoise	0.124	0.163	FireFlow	0.117	0.151
InfEdit	0.096	0.116	GNRi	0.038	0.043	FlowEdit	0.108	0.139
NPI	0.085	0.107				RF-Edit	0.103	0.128

Appendix

Table 9: dCLIP under the training encoder (ViT-L/14) and a held-out encoder (OpenCLIP ViT-bigG/14). Within-block ordering is identical.

Figure 10: Objective ablation on Z-Image Turbo over training iterations (PIE-Bench,

\omega{=}2.5

). LPIPS

{}_{\text{wh}}

(VGG-16)

\downarrow

, CLIP-Score

\uparrow

, dCLIP

\uparrow

, and ImageReward

\uparrow

for the full objective and variants that remove the LF term, the KL term, or the edit term ( No CLIP ), and that replace directional dCLIP with absolute CLIPScore. Without a directional edit term the model collapses toward the identity solution (low LPIPS, near-zero dCLIP and ImageReward); the full objective gives the best edit quality across training.

Figure 11: Prior-regularizer ablation on Z-Image Turbo: quality–preservation frontiers (PIE-Bench,

n{=}700

\omega

sweep). ImageReward

\uparrow

(left) and dCLIP

\uparrow

(right) vs. LPIPS

{}_{\text{wh}}

(VGG-16)

\downarrow

; upper-left is best, markers labelled with the guidance scale

\omega

. Removing either the KL or the low-frequency (LF) term shifts the frontier below and to the right of the full model, with No LF the weakest.

Figure 12: Adapter ablation over training iterations (PIE-Bench,

\omega{=}2.5

). LPIPS

{}_{\text{wh}}

(VGG-16)

\downarrow

, dCLIP

\uparrow

, and ImageReward

\uparrow

for three training regimes: only the forward adapter

F_{\phi}

, only the backward adapter

G_{\theta}

, and both jointly (full). The backward-only model plateaus at high LPIPS; the forward-only model is unstable and settles at lower dCLIP and ImageReward. Joint training of both adapters reaches markedly lower LPIPS while remaining competitive on edit quality, showing that inversion and generation must be calibrated jointly.

Figure 13: Editing quality–preservation frontiers (PIE-Bench,

n{=}700

): ImageReward

\uparrow

vs. LPIPS

{}_{\text{wh}}

(VGG-16)

\downarrow

, upper-left is best. Red curve sweeps our guidance scale

\omega

(marker labels); baselines are single points. Our frontier dominates the baselines on all three backbones (SD1.5, SDXL-Turbo, Z-Image Turbo).

Figure 14: Additional qualitative results on Z-Image Turbo (flow-based) against RF-Edit, FlowEdit, FireFlow, and UniEdit-Flow (one edit per row). Our method applies each edit while preserving structure, identity, and background.

Figure 15: Additional qualitative results on SDXL-Turbo against ReNoise, GNRi, and TurboEdit (one edit per row).

Figure 16: Additional qualitative results on SD1.5 (LCM-DreamShaper) against iCD, NPI, NTI, SwiftEdit, InfEdit, ChordEdit, and G&R (one edit per row).

Figure 17: Image reconstruction on Z-Image Turbo (flow-based) against RF-Edit, FireFlow, and UniEdit-Flow. Each method inverts the source image and reconstructs it under the source prompt with classifier-free guidance disabled; a faithful method returns the input. Our adapters recover fine detail, color, and structure most closely.

Figure 18: Image reconstruction on SDXL-Turbo against GNRi and ReNoise. Our method preserves identity and detail, whereas the baselines lose structure or drift in appearance.

Sweep	$\lambda$	CLIP ${}_{\text{wh}}\uparrow$	dCLIP $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	IR $\uparrow$
$\lambda_{\mathrm{LF}}$ ( $\lambda_{\mathrm{KL}}{=}0.1$ )	0	26.13	0.1423	0.2071	0.8171	21.49	0.6015
	0.25	26.66	0.1641	0.2067	0.8062	20.71	0.8302
	0.5	26.94	0.1677	0.2269	0.7922	20.21	0.8626
	1	26.44	0.1541	0.1959	0.8296	22.01	0.7383
	2	27.01	0.1714	0.2636	0.7586	18.93	0.8501
$\lambda_{\mathrm{KL}}$ ( $\lambda_{\mathrm{LF}}{=}1$ )	0	26.06	0.1350	0.1901	0.8306	21.81	0.6029

Appendix

Table 10: Loss-weight sweeps on Z-Image-Turbo (step 600, guided inference of Eq. 23 at

\omega{=}2.5

; internal scorer as in Table 1 ). Each row is a separately trained run; the

\lambda{=}0

rows and the bold defaults coincide with the corresponding rows of Table 1 .

Figure 19: Image reconstruction on SD1.5 against full-step DDIM inversion, iCD, and SwiftEdit. Our method reconstructs the input most faithfully, surpassing even full-step DDIM despite operating in the few-step regime.

Text-Guided Image Editing

Text-guided image editing changes selected image content from natural-language instructions while aiming to preserve everything else.

See all 166 papers

May 4, 2026

DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing

With recent advancements in large-scale pre-trained text-to-image (T2I) models, training-free image editing methods have demonstrated remarkable success. Typically, these methods involve adding noise to a clean image via an inversion process, followed by separate denoising steps for the reconstruction and editing paths during the forward process. However, since the reconstruction path is approximated using noisy latents from mismatched timesteps, existing methods inevitably suffer from accumulated drift, which fundamentally limits reconstruction fidelity. To address this challenge, we systematically analyze the inversion process within the flow transformer and propose DirectEdit, a simple yet effective editing method that eliminates the inherent reconstruction error without introducing additional neural function evaluations (NFEs). Unlike most prior works that attempt to rectify the inversion path, DirectEdit focuses on directly aligning the forward paths, enabling precise reconstruction and reliable feature sharing. Furthermore, we introduce a preservation mechanism based on attention feature injection and multi-branch mask-guided noise blending, which effectively balances fidelity and editability. Extensive experiments across diverse scenarios demonstrate that DirectEdit achieves efficient and accurate image editing, delivering superior performance that outperforms state-of-the-art methods. Code and examples are available at https://desongyang.github.io/Directedit.

Desong Yang, Mang Ye

National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China.

Flow Matching Text-Guided Image Editing Diffusion-Based Image Editing Image Editing

Ask PDF

May 31, 2026

Training-free image inversion for one-step diffusion models

In this work, we introduce a novel training-free inversion (TFinv) framework for one-step diffusion models,addressing key challenges in real image inversion and editing. We first identify two critical factors hamperingreal-image inversion and editing: (1) Initial Latent Editability, which is related to the distance between theinitial noise and the ideal Gaussian distribution, and (2) Caption Gap, which means the alignment betweentext captions and image representations. Both factors influence inversion efficiency and the editability ofone-step diffusion models. Then, we propose two novel techniques: iterative noise alignment (iterNA), whichminimizes the distribution gap to align with the normal Gaussian distribution, and suffix learning (suffL),which enhances text-to-image caption alignment by introducing learned suffix prompt tokens. These techniquesenable precise inversion of input images into their initial noise representations and facilitate image editing.Furthermore, we propose a mask-based editing technique for localized edits while preserving backgroundintegrity. Comprehensive experiments on the PIE-Bench dataset validate that our method TFinv not onlyachieves state-of-the-art performance in one-step diffusion editing, but also significantly outperforms existingmultistep approaches in efficiency. The code is available at https://github.com/tttao-uwu/TFinv.git.

Tao Wu, Senmao Li, Yaxing Wang +3

Computer Vision Center, Universitat Autonoma de Barcelona, Barcelona, 08193, Spain · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, United Arab Emirates · Jilin University, 130012, Changchun, Jilin, China +3

Diffusion Model Inversion Text-Guided Image Editing One-Step Generative Modeling Diffusion-Based Image Editing

Ask PDF

Aug 10, 2026

Diffusion Image Editing via Asynchronous Token Decoding

Text-guided diffusion image editing aims to modify semantic attributes of an image while preserving its identity, layout, and background. However, naïvely switching the text condition during sampling often causes global drift, as denoising dynamics propagate changes across tokens and can disrupt unedited regions. To address this issue, we propose Asynchronous Token Decoding Edit (ATDEdit), an inference-time framework that views each sampler step as a parallel update of a globally coupled token matrix and enables token-indexed condition switching with differentiated update policies. Instead of applying synchronous target-conditioned updates to all tokens, ATDEdit estimates editable locations using token-wise conditional surprisal and applies target-conditioned corrections to the selected token set. It supplies source key/value memory at keep-token positions and projects selected keep-token latent rows back to their source values; these operations promote background preservation but do not constitute a pixel-level invariance guarantee. This approach combines local editing and background preservation without external or user-provided spatial masks and without model fine-tuning. On PIE-Bench, ATDEdit achieves the strongest reported preservation metrics, including 27.44~dB PSNR and 0.055 LPIPS, while retaining competitive semantic alignment.

Yang Shi, Liangsi Lu, Minzhe Guo +4

Guangdong University of Technology Guangzhou, China · Hong Kong Baptist University Hong Kong, China · Peking University Beijing, China +1

Text-Guided Image Editing Diffusion-Based Image Editing Diffusion Models

Ask PDF

Abstract

Text-Guided Image Editing

Explore similar work

DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing

Training-free image inversion for one-step diffusion models

Diffusion Image Editing via Asynchronous Token Decoding