LOCI: Localized Continuous Control for Multi-Subject Image Editing

1 Introduction

• Continuous editors lack image-region binding, while localized editors apply edits at a fixed strength; neither family combines localization, continuous per-region control, and simultaneous subject control.The introduction identifies these capabilities as a gap in existing approaches.Source: §1 · pp. 1–2
• LOCI uses per-subject regions and instructions to apply independently chosen edit strengths while attention masking isolates edits from neighboring regions.Regions may be specified as bounding boxes or segmentation masks.Source: §1 · p. 2
• Because attention can spread a regional residual’s effect, LOCI trains LoRA under instance-aware attention masking so its edit directions become local and can be scaled independently.The introduction contrasts this with conditioning-path modulation, which does not linearly interpolate between unedited and edited outputs.Source: §1 · p. 2
• Regional LoRA replaces one image-wide strength coefficient with separate regional residual scaling, composing subject-specific edits while leaving the background unedited.The paper represents each edit with an instruction, spatial support, and continuous coefficient.Source: §1 · p. 3
• The authors curate 128 scenes and 427 source–region–edit tuples, then define four protocols for target fidelity, leakage, continuity, and slider independence.The scenes span people, animals, vehicles, furniture, and mixed compositions with 2–7 subjects.Source: §1 · p. 3

2 Related Work

• LOCI builds on Flux.1-Kontext, whose concatenated source-image latent tokens are a property the construction relies on.The passage identifies Black Forest Labs (2025) as the Flux.1-Kontext reference.Source: §2 · p. 3
• Attention manipulation and layout conditions localize edits, while masked inpainting uses binary regions and does not provide graded strength.Zaccagnino et al. (2026) use box-layout masks to block cross-instance leakage; the paper adopts this as its spatial backbone.Source: §2 · p. 3
• Concept Sliders, SliderEdit, and Kontinuous Kontext provide continuous controls, but SliderEdit and Kontinuous Kontext reach subjects through language rather than user-specified image regions.The paper cites Gandikota et al. (2023), Zarei et al. (2025), and Parihar et al. (2026).Source: §2 · p. 3
• Several multi-object and regional-adapter methods edit multiple subjects at fixed intensity, while the described approaches do not apply distinct per-subject edits or isolate regions at the attention level.The passage discusses Chakrabarty et al. (2024), Dalva et al. (2025), Foteinopoulou et al. (2026), Cai et al. (2026), Xiong et al. (2026), and Gholami and Xiao (2024).Source: §2 · p. 3
• LOCI unifies region-based addressing, per-token LoRA scaling, and instance-disentangled attention for independently controlled edits across several subjects in one forward pass.Source: §2 · p. 3

3 Method

• The method takes per-subject instructions, binary region masks, and scalar controls; a bounding box can be converted into a mask, with a corresponding box used for attention routing.Control α_n = 1 leaves a subject unedited, α_n = 0 applies the nominal edit, and α_n < 0 extrapolates beyond it.Source: §3 · pp. 3–4

3.1 Preliminaries

• LOCI uses a frozen Flux.1-Kontext backbone and partitions text, target-image, and context-image tokens by subject association.The target latent is denoised into the edited image, while the fixed context latent encodes the source image.Source: §3.1 · pp. 4–5

3.2 Regional LoRA

• Regional LoRA replaces the image-wide scalar adapter coefficient with a per-token vector, assigning region-specific strengths and zero adapter contribution to background tokens.The token-space region gate marks a token when any source pixel it covers belongs to that subject’s region.Source: §3.2 · p. 5

3.3 Instance-Disentangled Attention

• Instance-Disentangled Attention restricts subject tokens to their own instruction and image partitions in middle blocks, then permits cross-region visual reads in outer blocks for coherence.The same attention schedule is used in training and inference.Source: §3.3 · pp. 5–6

3.4 Single-Region Suppression Training

• Single-region suppression training uses the frozen editor’s neutral-instruction prediction as a target for one adapter trained on real-instruction examples.Both passes share the routing schedule, isolating edit suppression at the routing configuration used at inference.Source: §3.4 · p. 7

3.5 Multi-Region Inference

• At inference, one shared adapter supports N > 1 controls, including different signs for suppression and extrapolation, without retraining for a particular subject count.Each subject’s instruction and selected image tokens share its control coefficient; queries outside that subject cannot read its instruction block.Source: §3.5 · p. 7

4.1 Experimental Setup

• The benchmark contains 128 multi-subject scenes with 2–7 subjects per scene and 427 annotated subject–edit pairs.Scenes are synthetic, generated with Flux.2, and disjoint from HQ-Edit training data by construction; spatial annotations are supplied by the user at inference.Source: §4.1 · pp. 7–8

4.2 Quantitative Evaluation

• Five methods compare instruction-only control, routing, regional adaptation, and their combinations, including post-hoc composition versus routing-aware training.The component contrasts isolate routing alone, Regional LoRA alone, both together, and whether routing must be present during adapter training.Source: §4.2 · p. 8
• Protocol C tests whether multiple requested edits remain effective when composed in one image by comparing nominal-edit and neutral outputs.It scores each subject crop against its edit description and measures the similarity improvement from neutral output to nominal edit.Source: §4.2 · pp. 8–9
• Protocol D sweeps two subject controls on a 7 × 7 grid to distinguish each control’s intended response from cross-talk caused by the other.It evaluates two-subject scenes and applies only to slider-capable methods.Source: §4.2 · pp. 8–9
• Protocols A and B assess single-target and independent multi-control trajectories using extrapolation, disentanglement, direction correlation, and mean range.Extrapolation is maximum target similarity over the control trajectory; direction correlation measures ordering, and mean range measures the span of target similarity.Source: §4.2 · pp. 8–9

Implementation details.

• LOCI achieves the highest extrapolation and lowest disentanglement on every metric and both encoders across Protocols A and B.Its strongest preservation coincides with strongest target alignment, rather than resulting from weaker editing.Source: Implementation details · p. 9
• LOCI provides the most reliably ordered response and widest semantic range under both protocols and both encoders.Source: Implementation details · p. 9
• Under Protocol C, LOCI nearly matches IDAttn’s best coverage while retaining an independent slider per subject; IDAttn alone has no continuous control.The authors characterize the coverage cost of continuous control as small.Source: Implementation details · p. 9
• Under Protocol D, only routing-aware adapter training yields both the strongest intended response and lowest cross-talk on both measures.Regional adaptation without routing has the worst cross-talk, while post-hoc composition contains edits well but responds weakly; LOCI’s perceptual margin over post-hoc composition is small.Source: Implementation details · p. 9

4.3 Qualitative Results

• Figure 3 presents four multi-subject scenes, with rows sweeping one subject’s control, the other’s control, or both.Edit strength increases from left to right; the final block applies one instruction to both annotated regions.Source: §4.3 · p. 10

5 Conclusion

• LOCI edits multiple subjects in one image at independently chosen strengths within a single forward pass.It replaces a scalar adapter coefficient with a per-token vector aligned to the token partition and trains under the routing used at inference.Source: §5 · p. 10

A Extended Related Work

• Differential Diffusion and Layered Diffusion Brushes vary regional change or regeneration amount, but do not provide distinct subject edits with region isolation.Both rely on manually specified regions that can interact without isolation.Source: A Extended Related Work · p. 14
• Prior work documents cross-subject interference as semantic leakage in attention and text embeddings, motivating a dedicated isolation mechanism.Source: A Extended Related Work · p. 14

B Why We Do Not Report Semantic Continuity

• SliderEdit’s continuity score bins similarity values across their observed range, then compares bin counts with a uniform expectation.Source: B Why We Do Not Report Semantic… · p. 14
• Because the score ignores order and rescales the observed range, an oscillating or barely responsive trajectory can score highly.The paper gives 0.005 versus 0.30 as examples of ranges that are stretched to fill the same bins.Source: B Why We Do Not Report Semantic… · p. 14
• The reported diagnostics are Spearman correlation for trajectory order and mean range for trajectory scale.Source: B Why We Do Not Report Semantic… · p. 14

C Aggregation

• For most metrics, subject-level measurements are averaged within each source image before equal averaging across images, accounting for correlated pairs.Trajectory diagnostics instead use each source–region–instruction trajectory as the unit and average over all 427 trajectories.Source: C Aggregation · p. 15

• Regional success rate counts when an edit improves its target score, while worst-region gain exposes scenes where composition sacrifices even one subject.Source: D Formal Metric Definitions · p. 15
• For two controls, Protocol D evaluates a 7 × 7 grid and averages each region’s score change over 42 adjacent pairs per control direction.The pairs differ by one step in a single control.Source: D Formal Metric Definitions · p. 15
• Perceptual cross-talk uses LPIPS distances between regional crops of neighboring outputs; because absolute differences lose sign, mean signed response is checked separately.Both cross-talk terms compare neighboring generated outputs rather than comparing each output with the source.Source: D Formal Metric Definitions · p. 15

• The 128 evaluation scenes were curated from authored specifications and Flux.2 source candidates, retaining images with the intended subject count and regenerating failed scenes.Candidates were generated at 1024×1024 with 28 denoising steps, guidance scale 3.5, and typically four fixed seeds per scene.Source: E Benchmark Construction · p. 15
• Grounding-DINO detections were ordered left to right and passed to SAM2 to produce subject silhouettes, boxes, and masks for each scene.The stored annotations also include detection confidence, mask area, and ordered instructions.Source: E Benchmark Construction · p. 16
• Instructions give each subject one visually verifiable atomic edit, using SliderEdit’s prompt construction with LLM assistance and manual curation.Flux.2 generates the sources, while Flux.1-Kontext is the frozen editor at inference.Source: E Benchmark Construction · pp. 15–16

• The training manifest contains 22,482 examples spanning 21,962 instructions after filtering HQ-Edit data and removing masks smaller than 3% of the image.Examples requesting pose, movement, size, rotation, or viewpoint changes are also discarded; instructions are not resampled or synthesized.Source: F Implementation Details · p. 16
• Training updates only a rank-16 LoRA adapter without dropout; the VAE, text encoders, and scheduler remain frozen wherever applicable.The adapter is stored in fp32 while the frozen backbone computes in bf16.Source: F Implementation Details · p. 16
• The write gate projects each subject mask to a 64×64 token grid for 1024×1024 inputs, using cellwise maxima that preserve thin structures but can dilate gates by one token.Token cells span 16×16 pixels, and the resulting scale is applied to text and image entries in the corresponding transformer streams.Source: F Implementation Details · p. 16
• A single-region training example has 8,392 joint tokens: 200 positions per instruction block, plus 4,096 noise and 4,096 source-image tokens.Source: F Implementation Details · p. 16
• Both adapters train for 10,000 steps on 8×H200 GPUs with effective batch size 8, AdamW, and a cosine schedule with 500 warmup steps.The learning rate is 7 × 10−5, weight decay is 10−5, and gradient clipping is 1.0.Source: F Implementation Details · p. 16

• The vase targets are visually similar and differ mainly by position, the setting in which explicit spatial annotations matter most.Source: G Four-Subject Control · p. 16
• Figure 4 varies one chair’s control in each of the first four rows while holding the other three at α = 1, then moves all four controls together.The input shows four spatial annotations in red, blue, green, and orange; columns show edited outputs, with the unedited scene at left.Source: Figure 4 · p. 17
• Figure 5 varies one vase’s control per row before a final row moves all four together; its first column is the unedited output at α = 1.Source: Figure 5 · p. 18