MIRAGE: Benchmarking and Aligning Multi-Instance Image Editing

Ziqian Liu; Stephan Alaniz

cs.CVApr 6, 2026

MIRAGE: Benchmarking and Aligning Multi-Instance Image Editing

Authors: Ziqian Liu, Stephan Alaniz

Organizations: LTCI, Télécom Paris, Institut Polytechnique de Paris, France

Abstract

Instruction-guided image editing has seen remarkable progress with models like FLUX.2 and Qwen-Image-Edit, yet they still struggle with complex scenarios with multiple similar instances each requiring individual edits. We observe that state-of-the-art models suffer from severe over-editing and spatial misalignment when faced with multiple identical instances and composite instructions. To this end, we introduce a comprehensive benchmark specifically designed to evaluate fine-grained consistency in multi-instance and multi-instruction settings. To address the failures of existing methods observed in our benchmark, we propose Multi-Instance Regional Alignment via Guided Editing (MIRAGE), a training-free framework that enables precise, localized editing. By leveraging a vision-language model to parse complex instructions into regional subsets, MIRAGE employs a multi-branch parallel denoising strategy. This approach injects latent representations of target regions into the global representation space while maintaining background integrity through a reference trajectory. Extensive evaluations on MIRA-Bench and RefEdit-Bench demonstrate that our framework significantly outperforms existing methods in achieving precise instance-level modifications while preserving background consistency. Our benchmark and code are available at https://github.com/ZiqianLiu666/MIRAGE.

Figures & tables

Figure 1: Limitations of current state-of-the-art image editing models. Although these models successfully follow the editing instructions, noticeable limitations remain. (i) blue circles indicate failures in preserving local consistency within the target regions; (ii) green circles highlight unintended modifications in the background.

Dataset	Num. Image	Multi- Instance	Avg. Instances	Compositional Instructions	Avg. Instructions / Image	Referable
MagicBrush [ 46 ]	535	$\times$	1	$\times$	2	$\times$
HIVE [ 47 ]	200	$\times$	1	$\times$	5	$\times$
OmniEdit [ 41 ]	62	$\times$	1	$\times$	7	$\times$
ImgEdit [ 45 ]	811	Partial	2	Partial	4	$\checkmark$
RefEdit-Bench [ 28 ]	200	Partial	3	Partial	1	$\checkmark$
MIRA-Bench	100	$\checkmark$	4	$\checkmark$	5	$\checkmark$

Table 1: Comparison of image editing benchmarks. MIRA-Bench is the first benchmark to focus on multi-instance editing with 3-5 similar instances and 5 compositional edit instructions per image.

Figure 2: MIRA-Bench construction. (a) Images are generated from source prompts via FLUX.2 [ 4 ] . (b) A VLM extracts editing instructions and referring expressions. (c) VLM-predicted boxes are refined by SAM2 [ 32 ] into masks. Final benchmark triplets (image, instruction, mask) are human-validated.

Figure 3: Example images and instructions from MIRA-Bench. The multiple similar instances and compositional instructions pose a challenge to SOTA models which introduce unintended modification in these complex scenarios. Through MIRAGE, we achieve precise instance-level editing while preserving background consistency.

Figure 4: Overview of MIRAGE. (1) A VLM parses the global instruction into instance-level edits and localizes target regions via bounding boxes. (2) Regional branches perform parallel diffusion editing and inject their latents into the global branch; later time steps continue global editing under the global instruction while preserving background latents from the reference trajectory.

Figure 5: Qualitative results on MIRA-Bench. Integrating MIRAGE into SOTA models effectively mitigates over-editing in multi-instance scenarios while preserving detail consistency as much as possible.

Figure 6: Effect of latent replacement time step $\rho$ on MIRA-Bench . Results obtained with FLUX.2 [Dev]. Red circles highlight artifacts or over-editing regions.

Figure 7: Effect of target and background replacement strategies on MIRA-Bench. These results are obtained with FLUX.2 [Dev] using a replacement interval of [0.6T, T]. Red circles indicate over-editing and local detail inconsistencies.

Figure 8: Performance under increasing instruction complexity. Results on MIRA-Bench using a replacement interval of

[0.6T,\,T]

Method	PF $\uparrow$ avg@3	Cons $\uparrow$ avg@3	PQ ${}_{\mathrm{avg}}\uparrow$	Overall ${}_{\mathrm{avg}}\uparrow$	Parse (s)	Detect (s)	Inference (s)	Total Time $\downarrow$
Flux.2 [Klein]-base-9B (Sequential)	6.124	7.895	7.173	6.628	–	–	448.99	448.99
Flux.2 [Klein]-base-9B	6.046	8.646	8.988	7.372	–	–	90.13	90.13
Flux.2 [Klein]-base-9B + MIRAGE	7.709	8.796	8.836	8.253	2.90	8.37	74.59	85.87

Table 4: Runtime-performance comparison of different inference strategies on MIRA-Bench. Results obtained with FLUX.2 [Klein]-Base-9B under a replacement interval of [0.6T, T].

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : The Image descriptions and the corresponding generated image of MIRA-Bench.

Figure S2 : MIRA-Bench sample examples. The first row shows the synthesized original images, the second row presents the corresponding ground-truth (GT) masks of the target regions, and the third row displays the editing instructions constructed based on the generated image semantics and the source prompts.

Figure S3 : Additional qualitative results on MIRA-Bench.

Figure S4 : Sequential vs. joint editing inference. Applying instructions sequentially leads to accumulated artifacts and structural inconsistencies due to repeated image re-editing. In contrast, MIRAGE performs joint multi-region inference within a single denoising trajectory, preventing error accumulation and maintaining global coherence.

Figure S5 : Qualitative results on RefEdit-Bench . Across all compared methods, integrating our MIRAGE consistently improves performance, substantially outperforming both the baseline models and their original versions.

Figure S6 : Effect of latent replacement time-step selection on RefEdit-Bench. Results are obtained using FLUX.2 [Dev] under different latent replacement intervals.

Figure S7 : Effect of target and background replacement strategies on RefEdit-Bench. These results are obtained with FLUX.2 [Dev] using a replacement interval of $[0.8T,\,T]$ .

Figure S8 : Limitation of applying MIRAGE to Qwen-Image-Edit-2511 Although applying MIRAGE mitigates over-editing, the internal processing mechanism of Qwen still leads to degraded detail resolution in the edited target regions in certain cases (as highlighted by the red circles), resulting in blurred textures and indistinct structures. This effect is particularly pronounced in removal-based edits.

Figure S9 : Prompts used for instruction parsing and referring-expression localization prior. Left: prompt for the information-extraction-based parser, which decomposes complex editing instructions

I

into aligned referring objects

r_{k}

and edit sub-instruction

I^{k}

. Right: prompt for the referring-expression localizer, which grounds each referring expression

r_{k}

to a spatial region

x_{k}

before image editing.

Figure S10 : Prompts used for image description in MIRA-Bench construction. Top: Prompt for generating (repeated instance category, scene) pairs. Middle: Prompt for generating image descriptions. Bottom: Prompt for verifying and refining the image descriptions.

Figure S11 : Prompts used for instruction and refer-object generation in MIRA-Bench construction. Top: Prompt for extracting instance-level details from the image description. Middle: Prompt for generating editing instructions. Bottom: Prompt for generating referring object expressions.

Figure S12 : Evaluation prompt of PF and Cons score

Figure S13 : Evaluation prompt of PQ score

Image Editing

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

See all 315 papers

Editing Everything Everywhere All at Once
Jun 30, 2026Fabio Quattrini, Carmine Zaccagnino, Enis Simsar +4Text-Guided Image Editing Attention Control in Diffusion Models
University of Modena and Reggio Emilia · ETH Zurich · Google
Borrowing from anything: A generalizable framework for reference-guided instance editing
Dec 17, 2025Shengxiao Zhou, Chenghua Li, Jianhao Huang +2Disentangled Representation Learning
University of Chinese Academy of Sciences, China · Nanjing Forestry University, Nanjing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China +1
EditCaption: Human-Refined SFT and HAE-DPO for Image Editing Instruction Synthesis
Apr 9, 2026Xiangyuan Wang, Honghao Cai, Yunhao Bai +9Direct Preference Optimization Image Captioning
Peking University · Xiaohongshu Inc. · The Chinese University of Hong Kong, Shenzhen +2

Jun 30, 2026J/K move · Enter open · S save

Editing Everything Everywhere All at Once

Fabio Quattrini, Carmine Zaccagnino, Enis Simsar, Marta Tintoré Gazulla+3

University of Modena and Reggio Emilia · ETH Zurich · Google

Text-Guided Image Editing Attention Control in Diffusion Models Diffusion-Based Image Editing Image Editing Diffusion Models Multimodal Diffusion Models

Open paper AskPDF arXiv

Editing multiple elements of an image in a single forward pass is a practical alternative to multi-turn image manipulation, offering improved efficiency and potentially better harmonization. However, when several instructions target different regions, semantic interference often leads to attribute leakage and poor edit disentanglement, especially as the number of edits increases. In this work, we propose MICE (Multi-Instance Concurrent Editing), a training-free strategy for scalable multi-instance image editing with Multimodal Diffusion Transformers. MICE modifies the additive bias of joint attention to regulate interactions between instance-specific edit instructions, latent, and context tokens identified via user-provided segmentation masks. Specifically, MICE allows intra-instance attention, penalizes interactions between neighboring region tokens, and suppresses unrelated cross-instance attention. As a result, our method enforces attribute binding while preserving global visual consistency. We evaluate MICE on LoMOE-Bench and introduce MICE-Bench, a more challenging benchmark with an average of 8.5 concurrent edits per image. The experiments demonstrate that our approach outperforms strong baselines and recent competitors in terms of visual quality preservation and faithfulness to the editing instructions.

Abstract

Image Editing

Explore similar work

Editing Everything Everywhere All at Once

Borrowing from anything: A generalizable framework for reference-guided instance editing

EditCaption: Human-Refined SFT and HAE-DPO for Image Editing Instruction Synthesis