Inference-time Trajectory Optimization for Structure-Preserving Manga Image Editing

Ryosuke Furuta

cs.CVMar 29, 2026

Inference-time Trajectory Optimization for Structure-Preserving Manga Image Editing

Authors: Ryosuke Furuta

Organizations: Mantra Inc.

Abstract

We present a lightweight, training-free trajectory correction method that adapts a pretrained image editing model to each input manga image using only the input itself. Despite recent progress in pretrained image editing, such models often underperform on manga because they are trained predominantly on natural-image data, while re-training or fine-tuning them on manga is costly and raises copyright concerns. Many manga image editing tasks encountered in practice are structure-preserving, requiring local details to be modified while the input's global composition is retained. To support this common editing setting, our method corrects the early editing trajectory by anchoring it to an empty-prompt reconstruction trajectory. Experiments indicate improved performance in the main text-removal setting, while qualitative examples suggest better composition preservation in screentone synthesis. With FLUX.1 Kontext on an RTX A6000, the method incurs 11% runtime overhead and 0.1% peak-memory overhead; an additional runtime measurement with Qwen Image Edit 2509 on an NVIDIA H200 shows only a 0.1% increase.

Figures & tables

Panel of Figure 1 — Figure 1: Editing results with an empty prompt. ©Kato Masaki

Panel of Figure 2 — Figure 2: Examples of manga image editing. In each pair, the left image is the input and the right image is the ground truth. ©Kato Masaki ©Yabuno Tenya, Watanabe Tatsuya

Method	Full image			Text region
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline	11.84	48.02	41.13	8.71	28.90
Ours	22.27	86.25	7.32	18.50	78.13

Table 1: Evaluation of generation results under an empty prompt

Method	Model	Full image			Text region
Method	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline	Qwen Image Edit 2509	19.62	72.76	7.16	19.07	60.77
Noise Inversion	Qwen Image Edit 2509	20.29	75.66	7.23	18.94	61.80
Ours	Qwen Image Edit 2509	22.07	80.61	6.00	20.04	64.15
FlowChef ( Patel et al. 2025 )	Qwen Image 2512	13.97	52.11	39.68	5.68	8.42
FlowEdit ( Kulikov et al. 2025 )	Qwen Image 2512	10.66	35.24	10.69	8.96	16.57

Table 2: Quantitative comparison for text removal

Panel of Figure 3 — Figure 3: Qualitative comparison for text removal. ©Nagano Noriko ©Omi Ayuko

Method	Full image			Text region
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline	19.62	72.76	7.16	19.07	60.77
Edit prompt	21.57	80.68	6.74	18.69	61.99
Empty prompt	22.07	80.61	6.00	20.04	64.15

Table 3: Effect of the prompt used for trajectory correction

Panel of Figure 4 — Figure 4: Effect of the prompt used for trajectory correction. ©Okuda Momoko

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Full image			Text region
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline	19.62	72.76	7.16	19.07	60.77
Straight path	19.69	73.13	7.74	19.13	61.12
Optimized path	22.07	80.61	6.00	20.04	64.15

Appendix

Table 4: Effect of the optimized trajectory

Practical operating range
Method	Hyper-	Full image			Text region
Method	param.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline		19.62	72.76	7.16	19.07	60.77
Ours	$M=1$	20.89	76.87	6.94	19.71	62.56
	$M=3$	22.07	80.61	6.00	20.04	64.15
	$M=5$	22.31	81.67	6.01	19.82	63.59

Appendix

Table 5: Stability within practical ranges and interpretation using deliberately extreme hyperparameter settings

Method	Model	Full image			Text region
Method	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Baseline	Qwen Image Edit 2509	19.62	72.76	7.16	19.07	60.77
Ours	Qwen Image Edit 2509	22.07	80.61	6.00	20.04	64.15
Baseline	Qwen Image Edit 2511	21.19	82.63	10.16	18.51	62.86
Ours	Qwen Image Edit 2511	21.39	83.05	9.84	18.54	62.55
Baseline	FLUX.1 Kontext [dev]	26.42	90.47	6.60	21.00	66.97

Appendix

Table 6: Evaluation across image editing models for text removal

Method	Model	FID $\downarrow$	CMMD $\downarrow$
Baseline	Qwen Image Edit 2511	36.08	0.3140
Noise Inversion	Qwen Image Edit 2511	36.71	0.3123
Ours	Qwen Image Edit 2511	36.07	0.3129
FlowChef	Qwen Image 2512	81.88	0.9993
FlowEdit	Qwen Image 2512	61.32	0.4669
Baseline	FLUX.1 Kontext [dev]	92.97	0.3746

Appendix

Table 7: Quantitative evaluation of screentone synthesis

Panel of Figure 5 — Figure 5: Qualitative comparison for screentone synthesis. ©Saki Kaori ©Kurita Riku

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 298 papers

ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning
Oct 15, 2026Sara Dorfman, Nir Goren, Gal Barak +2Multi-Turn Image Editing Image Editing
Tel Aviv University
From Scale to Speed: Adaptive Test-Time Scaling for Image Editing
Feb 24, 2026Xiangyan Qu, Zhenlong Yuan, Jing Tang +9Test-Time Scaling Text-Guided Image Editing
Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences · AMAP, Alibaba Group +1
Scalable Minimal-Change Learning for Controllable Image Editing
Oct 5, 2026Shuo Chen, Fengming Huang, Yu Yao +2Text-Guided Image Editing Image Editing
Sydney AI Centre, The University of Sydney

Oct 15, 2026J/K move · Enter open · S save

ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning

Sara Dorfman, Nir Goren, Gal Barak, Daniel Cohen-Or+1

Tel Aviv University

Multi-Turn Image Editing Image Editing Text-Guided Image Editing

Open paper AskPDF arXiv

Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating a gap between the input distributions encountered during the editor's original training and multi-turn inference. We introduce a training scheme that makes multi-turn image editing stable under this accumulating distribution shift. A key is to force the editor to operate on its own outputs from previous editing turns during training, bringing the principle of Self-Forcing to image editing. The main challenge is that self-generated editing trajectories deviate from ground-truth trajectories, making direct supervision unreliable. We address this challenge through a distillation objective that leverages a teacher image editor to provide a reliable training signal when the model operates on its own previous outputs. We introduce an asymmetric conditioning scheme: while the student performs edits sequentially on its own outputs, the teacher remains anchored to the original clean image and receives a compound editing instruction. Experiments show substantially reduced degradation over long multi-turn sequences while maintaining editing quality and instruction adherence.

Abstract

Image Editing

Explore similar work

ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning

From Scale to Speed: Adaptive Test-Time Scaling for Image Editing

Scalable Minimal-Change Learning for Controllable Image Editing