LayerVerse: Finding the Sweet Spot
for KV-Injection in Training-Free Image Editing

ECCV 2026

Mikhail Zhirnov^1,2,3, Arsen Kuzhamuratov^1,2, Andrey Kuznetsov^1,2, Ivan Oseledets^2,3, Konstantin Sobolev^1,2,4

¹ FusionBrain Lab ² AXXX ³ Applied AI Institute ⁴ Lomonosov Moscow State University

Corresponding author: zhirnov.m.d@gmail.com

Paper Code Poster Video BibTeX

Car → motorcycle — Source image — Car → motorcycle

Change material to fabric — Source image — Change material to fabric

Delete rainbow — Source image — Delete rainbow

LayerVerse jointly selects block identity and injection scope: masked background protection and sparse global structural anchors.

Abstract

For image editing with Multimodal Diffusion Transformers (MM-DiT), training-free methods face a critical trade-off: precise background preservation limits editability, while high prompt fidelity degrades the original scene and object structure. We argue that this tension stems from suboptimal Key-Value (KV) injection strategies. Global KV-injection rigidly over-preserves the source, whereas strictly masked injection acts as localized inpainting, destroying the edited object’s structural identity.

We introduce LayerVerse, an optimization framework that resolves this fundamental dilemma of layer selection for KV-injection by assigning specific layers to two distinct roles: Masked Injection to precisely protect the background, and Global Injection to anchor the edited object’s structure. Formulating optimal layer allocation as a combinatorial problem, we model the editing error based on pairwise layer interactions and solve it globally via Mixed-Integer Linear Programming (MILP). Supported by lightweight, single-step automatic masking, LayerVerse seamlessly integrates into training-free, inversion-based editors, achieving highly competitive results on the PIE-Bench benchmark.

Method

Motivation: the KV-Injection Dilemma

Reusing the source image’s attention keys and values (KV) helps preserve its content. The challenge is to retain the background and object structure while allowing the requested edit.

Global injection collapse. Source features injected across the image can dominate generation and pull the result toward the original. Even a few global injections can prevent a meaningful edit.
Masked inpainting pitfall. Background-only injection preserves unedited pixels but withholds source features from the editing region. The object may be synthesized from scratch, losing its shape, pose, or identity.

Which blocks should receive source features — and should the injection be masked or global?

LayerVerse Finds the Right Role for Each Block

Jointly select block identity and injection scope: inactive, masked, or global.

Masked blocks preserve the background; sparse global blocks anchor object structure.

Modeling the editing error

Individual block effects are not independent. LayerVerse models the editing error with a second-order surrogate that includes pairwise interactions.

E(x) is approximately C plus the sum of w_i x_i and the sum of W_ij x_i x_j for i less than j.

$w_i$ : individual block effect $W_{i,j}$ : pairwise synergy or conflict

The coefficients are estimated on a hold-out dataset using no-injection, single-block, and block-pair evaluations.

Optimal role assignment with MILP

Combine the calibrated background and structure objectives under fixed block budgets. Each block is inactive, masked, or global.

Minimize background error E_bg(m+s) plus lambda times structure error E_str(s), with N active blocks and N_str global anchors.

$m_i$ : masked injection $s_i$ : global injection

Mixed-Integer Linear Programming finds the globally optimal assignment for the calibrated surrogate, without exhaustive search.

Auxiliary Automatic Masking for masked KV-injection

One-time architecture calibration. The resulting backbone-specific allocation is reused across editing solvers without recalibration.

Block Allocation and Calibration Details

Optimized block allocation

Backbone / budget	Masked KV0-based indices	Global KV0-based indices
FLUX.1-[dev]N = 8 · N_str = 1	Double `[1]` · Single `[32, 33, 34, 35, 36, 37]`	Double `[2]`
SD3.5-mediumN = 13 · N_str = 4	`[0, 1, 3, 4, 12, 14, 17, 18, 19]`	`[20, 21, 22, 23]`

Calibration dataset

115 held-out MS-COCO images. Error measurements use a standard Euler solver. The same subset is used to select masking parameters offline by maximizing IoU against ground-truth masks.

Computational cost

FLUX.1-[dev] · H100 · one-time offline calibration

Error evaluation · 115 images: ∼348 GPU-hours
Error evaluation · 10 images: ∼30 GPU-hours
Subsequent MILP solve: ∼1.97 seconds

Qualitative Results

Comparisons on PIE-Bench using FLUX.1-[dev] and SD3.5-medium.

Scroll sideways to see all three images, or switch to Before / after.

Source image

FireFlow

FireFlow + LayerVerse

Quantitative Evaluation on PIE-Bench

Lowest Average Rank: 2.88 on FLUX.1-[dev] · 2.44 on SD3.5-medium

Scroll horizontally to view all metrics.

Quantitative Evaluation on PIE-Bench · FLUX.1-[dev]

Method	LV	Structure	Background Preservation				CLIP Similarity		User Alignment		Rank
Method	LV	Distance ↓	PSNR ↑	LPIPS ↓	MSE ↓	SSIM ↑	Whole ↑	Edited ↑	HPSv3 ↑	EditScore ↑	Avg. ↓
RF-inversion	×	42.00	20.15	177.11	141.22	70.64	24.57	21.75	6.36	4.16	5.88
FlowEdit	×	27.96	21.92	112.99	95.68	83.35	25.41	22.29	6.81	4.99	4.13
DNAEdit	×	17.00	25.18	86.78	48.36	87.19	24.78	21.77	5.80	4.68	4.50
UniEdit-Flow	×	8.82	29.74	55.35	18.16	91.12	24.59	21.29	4.27	4.73	3.50
RF-Solver	×	11.87	27.81	73.21	28.17	88.35	23.74	20.53	4.93	3.71	5.88
RF-Solver	✓	9.45	30.16	37.62	18.01	92.66	24.30	21.23	5.16	4.93	3.38
FireFlow	×	9.48	28.33	64.59	24.51	89.23	23.68	20.45	4.88	3.60	5.88
FireFlow	✓	9.36	30.15	37.82	18.05	92.61	24.39	21.30	5.07	5.05	2.88

LV: LayerVerse. Best results are bold; second-best results are underlined. ↓ / ↑ indicates lower / higher is better.

Quantitative Evaluation on PIE-Bench · SD3.5-medium

Method	LV	Structure	Background Preservation				CLIP Similarity		User Alignment		Rank
Method	LV	Distance ↓	PSNR ↑	LPIPS ↓	MSE ↓	SSIM ↑	Whole ↑	Edited ↑	HPSv3 ↑	EditScore ↑	Avg. ↓
FTEdit	×	21.65	23.45	92.85	62.38	85.84	24.47	21.13	4.50	4.33	5.50
FlowEdit	×	23.16	23.24	93.58	69.67	85.10	26.59	23.15	6.47	5.92	3.50
FSI-Edit	×	14.02	26.60	84.43	36.79	86.31	25.49	22.01	5.66	4.82	3.63
DNAEdit	×	13.64	26.65	74.50	32.64	88.91	25.44	22.19	5.41	5.16	2.50
RF-Solver	✓	13.81	26.50	51.46	36.63	89.60	25.14	22.24	4.99	4.91	3.44
FireFlow	✓	13.52	26.58	50.85	35.95	89.70	25.11	22.25	5.01	4.93	2.44

LV: LayerVerse. Best results are bold; second-best results are underlined. ↓ / ↑ indicates lower / higher is better.

Ablations

Both Injection Roles Are Necessary

Global-only KV over-constrains the edit. Masked-only KV preserves the background but loses object structure. LayerVerse combines both roles for the best structure–editability balance.

Scroll sideways to compare all injection policies.

Source image

FireFlow

Only global KV

Only masked KV

FireFlow + LayerVerse

Visual ablation of KV-injection strategies.

Quantitative Comparison of Injection Roles

Injection Type	Structure	CLIP Similarity		User Alignment
Injection Type	Distance ↓	Whole ↑	Edited ↑	HPSv3 ↑	EditScore ↑
Standard FireFlow	9.48	23.68	20.45	4.88	3.59
Only Global KV	11.01	24.28	20.99	4.69	4.23
Only Masked KV	19.03	25.03	22.06	5.64	5.91
LayerVerse (Combined)	9.36	24.39	21.30	5.09	5.05

Pairwise Terms Are Necessary

The quadratic model outperforms the independent linear model in reconstruction quality at every tested budget (N = 2–15). Block synergies and conflicts make joint selection necessary.

Quadratic landscape analysis.

Calibration Data Sensitivity and Computational Cost

The structural anchor and core masked blocks remain stable under resampling. The imputed average-rank analysis shows competitive performance from b ≥ 10 calibration images, reducing offline error-evaluation cost from approximately 348 to 30 H100 GPU-hours.

Data sensitivity analysis: block stability, swap distance, and imputed average rank.

Data Sensitivity on PIE-Bench

Method	Structure	Background Preservation				CLIP Similarity		User Alignment		Rank
Method	Distance ↓	PSNR ↑	LPIPS ↓	MSE ↓	SSIM ↑	Whole ↑	Edited ↑	HPSv3 ↑	EditScore ↑	Avg. ↓
UniEdit-Flow	8.82	29.74	55.35	18.16	91.12	24.59	21.29	4.27	4.73	3.50
FireFlow + LV	9.21±0.37	30.29±0.21	36.39±2.08	17.64±0.56	92.77±0.19	24.39±0.01	21.28±0.03	5.10±0.03	5.01±0.06	2.88±0.00

± is the maximum metric deviation across the bootstrap candidate pool covering ≥99.9% of empirical mass.

BibTeX

@inproceedings{zhirnov2026layerverse,
  title={LayerVerse: Finding the Sweet Spot for KV-Injection in Training-Free Image Editing},
  author={Zhirnov, Mikhail and Kuzhamuratov, Arsen and Kuznetsov, Andrey and Oseledets, Ivan and Sobolev, Konstantin},
  booktitle={European Conference on Computer Vision},
  pages={96--113},
  year={2026},
  organization={Springer}
}

LayerVerse: Finding the Sweet Spot for KV-Injection in Training-Free Image Editing