Recent instruction-following image editors such as Flux.1-Kontext enable text-driven editing but lack fine-grained control over where and how much an edit is applied. Existing approaches provide either continuous but global control or spatially localized editing at a fixed strength. To our knowledge, none jointly supports explicit spatial localization, continuous per-region strength, and independent simultaneous control of multiple subjects. We present LOCI (LOcalized Continuous control for multi- subject Image editing). Given a multi-subject image with per-subject regions and instructions, LOCI edits each subject at an independently chosen, continuously adjustable strength, for instance gradually aging one person while adding a beard to another with minimal cross-region leakage. LOCI pairs subject-level attention binding for spatial isolation with a Regional LoRA layer that replaces the LoRA's scalar coefficient with a per-token scaling vector aligned to the spatial layout. A single adapter is trained at one suppression point, and continuous edit strengths follow by rescaling its regional residual at inference. Experiments on multi-subject scenes demonstrate that LOCI achieves strong edit fidelity with significantly reduced cross-region interference compared to baselines, while maintaining continuous control over edit strength
Figures & tables
Figure 1: LOCI gives every subject its own continuous edit control. From one input, the top row strengthens the edit on the first teapot while the second stays pinned at the neutral anchor, the middle row does the converse, and the bottom row moves both controls together. The boxes on the input are the spatial annotations supplied to the model: red for subject 1 and blue for subject 2, matching the row labels.
Figure 2: Architecture of LOCI. Stage 1 : the user supplies a source image, per-region bounding boxes Bn , edit instructions, and independent scalar controls αn . Stage 2 : text and image tokens are organized into the partition {Tn,Ln,Cn} indexed by subject, with trainable low-rank adapters ( ΔW , marked with ) whose residuals are gated per token by αn ; background tokens Lu,Cu receive zero adapter contribution. Stage 3 : the frozen Flux.1-Kontext transformer processes the sequence through 19 dual-stream and 38 single-stream blocks under a three-zone attention schedule, with harmonization in the outer blocks (0–9, 47–56) and disentanglement in the middle blocks (10–46). Inset : the query–key attention matrices for both regimes. Under disentanglement, each subject’s tokens attend only within their own partition; under harmonization, subject image tokens additionally read all visual tokens (dashed cells), restoring the cross-region evidence needed for background coherence.
Figure 3: Independent and joint control across four scenes. Each block shows its input and instructions on the left, with red and blue boxes marking the spatial annotations given to the model. The three rows sweep subject 1, subject 2, and both, and within a row the edit strengthens from left to right. see section 4.3 .
Extrapolation ↑
Direction corr. ↓
Mean range ↑
Disentanglement ↓
Protocol
Method
CLIP
SigLIP
CLIP
SigLIP
CLIP
SigLIP
LPIPS-Alex
LPIPS-VGG
DINOv2
Identity
A
(1) SliderEdit
0.2431
0.1114
−0.4907
−0.4666
0.0575
0.0470
0.3315
0.3826
0.1612
0.3640
A
(2) SliderEdit + IDAttn
0.2500
0.1187
−0.6326
−0.6192
0.0726
0.0587
0.1508
0.2010
0.0609
0.2598
A
(4) Regional LoRA without IDAttn
0.2445
0.1133
−0.5829
−0.5689
0.0761
0.0625
0.3085
0.3581
0.1700
0.3406
A
(5) LOCI (ours)
0.2527
0.1207
−0.7081
−0.6953
0.0894
0.0721
0.1467
0.1979
0.0550
0.2548
B
(1) SliderEdit
0.2426
0.1121
−0.5503
−0.5106
0.0589
0.0488
0.3346
0.3875
0.1471
0.3502
Table 1: Protocols A and B. Best per column bold , second best underlined , ranked within each protocol block.
Protocol C: joint editing
Protocol D: two-dimensional control
Method
Regional success ↑
Worst-region Δ CLIP ↑
Intended response ↑
Semantic cross-talk ↓
LPIPS cross-talk ↓
(1) SliderEdit
0.8231
0.0045
0.0307
0.6456
0.6597
(2) SliderEdit + IDAttn
0.8773
0.0122
0.0353
0.2776
0.1909
(3) IDAttn only
0.9153
0.0282
not slider-capable
(4) Regional LoRA without IDAttn
0.7577
−0.0015
0.0393
0.7146
0.9599
(5) LOCI (ours)
0.9099
0.0251
0.0458
0.2211
0.1860
Table 2: Protocols C and D. Protocol D applies only to slider-capable methods. Best per column bold , second best underlined .
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Four independent controls in one scene. The input carries four spatial annotations, drawn in red , blue , green and orange , each with its own instruction. Each of the first four rows moves one chair’s control while the other three are held at the neutral anchor α=1 ; the last row moves all four together. Every column shows an edited output; the unedited scene is the input on the left.
Figure 5: Four independent controls on a second scene. Each of the first four rows moves one vase’s control while the other three are held at the neutral anchor α=1 , and the last row moves all four together. The first column is the unedited output at α=1 .
Instruction-based image editing (IIE) aims to modify an image according to a natural language instruction. Despite recent advances in diffusion transformers, existing methods often introduce unintended changes to regions unrelated to the requested edit. We attribute this limitation to the absence of an explicit mechanism for edit localization. Different editing operations (e.g., subject addition, removal, and replacement) induce distinct spatial patterns, yet existing IIE models typically perform localization in a task-agnostic manner. To address this limitation, we propose a training-free, task-aware edit localization framework that exploits the intrinsic source and target image streams of IIE models. For each image stream, we construct feature centroids from attention-based edit cues, and then partition tokens into edit and non-edit regions based on feature similarity. Observing that effective localization is inherently task-dependent, we introduce a unified mask construction strategy that selectively leverages the source and target streams according to the editing task. We also provide a systematic analysis of our underlying insights and design choices. Extensive experiments on EdiVal-Bench demonstrate that our framework consistently improves content consistency in non-edit regions while maintaining strong instruction-following performance on top of Step1X-Edit and Qwen-Image-Edit.
Image editing should change only the attributes specified by an instruction while preserving everything else, yet current methods often make unintended changes. We treat this minimal-change principle as an optimization objective for instruction-based editing. Latent L1 regularization is a poor proxy for output locality in modern nonlinear generators and often requires supervision unavailable at scale. We instead optimize edit outcomes with reinforcement learning. An agentic vision-language reward model audits each source image, instruction, and edited image for two failure types: unimplemented requested changes and unintended changes. A group-level rubric merges and verifies these issues to provide consistent rewards across candidate edits without per-instruction human annotations. On FLUX.1 Kontext-dev, ARRO raises average EditScore from 5.21 to 5.88 across MinEval, MagicBrush, AnyBench, and Emu-Edit. On 600 evaluation examples, it reduces off-target pixel change by 8.4% relative to the base editor. Reward and SFT controls, blinded human evaluations, and transfer to OmniGen2 provide complementary evidence. Code: https://github.com/Showwwwwwwww/ARRO
Instruction-based image editing aims to modify source content according to textual instructions. However, existing methods built upon flow matching often struggle to maintain consistency in non-edited regions due to denoising-induced reconstruction errors that cause drift in preserved content. Moreover, they typically lack fine-grained control over edit strength. To address these limitations, we propose VeloEdit, a training-free method that enables highly consistent and continuously controllable editing. VeloEdit dynamically identifies editing regions by quantifying the discrepancy between the velocity fields responsible for preserving source content and those driving the desired edits. Based on this partition, we enforce consistency in preservation regions by substituting the editing velocity with the source-restoring velocity, while enabling continuous modulation of edit intensity in target regions via velocity interpolation. Unlike prior works that rely on complex attention manipulation or auxiliary trainable modules, VeloEdit operates directly on the velocity fields. Extensive experiments on Flux.1 Kontext and Qwen-Image-Edit demonstrate that VeloEdit improves visual consistency and editing continuity with negligible additional computational cost. Code is available at https://github.com/xmulzq/VeloEdit.