Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision

Jiacheng Chen; Songze Li; Han Fu; Baoquan Zhao; Wei Liu; Yanyan Liang; Li Qing; Xudong Mao

Source-linked AI summary

Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision

Jiacheng Chen, Songze Li, Han Fu, Baoquan Zhao, Wei Liu, Yanyan Liang, Li Qing, Xudong Mao

arXiv:2605.07940v1May 8, 2026cs.CV

TL;DR

Exemplar-based image editing typically requires two aligned image pairs, making scalable training-data curation difficult. Delta-Adapter instead learns a semantic transformation delta from one exemplar pair and injects it into an editing model, achieving stronger seen- and unseen-task performance than four baselines.

Problem
Existing exemplar-based editing methods require two aligned image pairs, making training-data curation difficult and limiting scalable model training.
Method
Delta-Adapter conditions a pre-trained editing model on a vision-encoder semantic delta from one exemplar pair, without textual guidance or directly exposing the edited image.
Results
On unseen tasks, Delta-Adapter achieves a GPT-A score of 4.008 versus 2.884 for RelationAdapter, while outperforming baselines overall on seen and unseen tasks.
Takeaways & Limitations
Single-pair supervision supports training from existing large-scale editing datasets and enables test-time adaptation for challenging unseen exemplars.
Takeaways & Limitations
The semantic-delta consistency loss relies on recovered denoised latents being sufficiently sharp for reliable feature extraction at the first denoising step.

Abstract

from arXiv · show

Exemplar-based image editing applies a transformation defined by a source-target image pair to a new query image. Existing methods rely on a pair-of-pairs supervision paradigm, requiring two image pairs sharing the same edit semantics to learn the target transformation. This constraint makes training data difficult to curate at scale and limits generalization across diverse edit types. We propose Delta-Adapter, a method that learns transferable editing semantics under single-pair supervision, requiring no textual guidance. Rather than directly exposing the exemplar pair to the model, we leverage a pre-trained vision encoder to extract a semantic delta that encodes the visual transformation between the two images. This semantic delta is injected into a pre-trained image editing model via a Perceiver-based adapter. Since the target image is never directly visible to the model, it can serve as the prediction target, enabling single-pair supervision without requiring additional exemplar pairs. This formulation allows us to leverage existing large-scale editing datasets for training. To further promote faithful transformation transfer, we introduce a semantic delta consistency loss that aligns the semantic change of the generated output with the ground-truth semantic delta extracted from the exemplar pair. Extensive experiments demonstrate that Delta-Adapter consistently improves both editing accuracy and content consistency over four strong baselines on seen editing tasks, while also generalizing more effectively to unseen editing tasks. Code will be available at https://delta-adapter.github.io.

1 Introduction

Delta-Adapter addresses the limitations of pair-of-pairs exemplar-based editing by learning transferable editing semantics from a single image pair without textual guidance. It extracts a semantic delta with a pre-trained vision encoder, injects it through a Perceiver-based adapter, and improves editing accuracy, content consistency, and generalization across editing tasks.

Motivation: Exemplar-based editing represents transformations through source/target image pairs, avoiding the difficulty of precisely articulating subtle appearance shifts and edit extent in words.This paradigm is also known as image analogy.
Problem: Existing methods use pair-of-pairs supervision, requiring two closely matched image pairs sharing edit semantics to predict the target of a new query pair.Uncontrolled discrepancies between pairs can make the intended edit difficult to isolate.
Method: Delta-Adapter learns transferable editing semantics under single-pair supervision without textual guidance by encoding the visual transformation as a semantic delta.The delta is computed from one exemplar pair using a pre-trained vision encoder and injected into a pre-trained image editing model through a Perceiver-based adapter.
Practical advantages: Training on individual source/target pairs enables direct use of existing large-scale editing datasets, broadening edit diversity and improving generalization to unseen edits.The single-pair formulation also enables efficient test-time adaptation for challenging unseen exemplars.
Results: Across diverse editing tasks, Delta-Adapter outperforms four strong baselines in editing accuracy, content consistency, and generalization to unseen edits.Test-time adaptation further substantially improves performance on unseen tasks.

2 Related Work

Related work spans diffusion-based image editing with diverse conditioning signals and exemplar-based methods that learn transformations from image pairs. Exemplar-based approaches include optimization-based methods that adapt parameters to encode pair-defined transformations.

Diffusion-based image editing: Diffusion models dominate high-quality image generation and editing, using text, masks, structure, and reference images as conditioning signals.Text provides flexible semantic control, while masks localize edits and geometric cues enforce spatial faithfulness.
Diffusion-based image editing: Reference-guided methods transfer appearance, identity, or style from a reference image to the target.They provide a complementary conditioning strategy to localization- and structure-focused editing methods.
Exemplar-based image editing: Exemplar-based editing learns visual transformations from image pairs, extending the classical image analogy framework into large generative models.Recent diffusion-based approaches are categorized by how they use the exemplar pair at test time.
Exemplar-based image editing: Optimization-based exemplar methods adapt learnable parameters to encode the transformation defined by the image pair.This is one broad category of diffusion-based exemplar editing approaches distinguished by test-time exemplar-pair usage.

3 Preliminary

The method builds on FLUX, which models image generation as rectified flow in latent space. Rectified flow linearly interpolates clean and noise latents while training a velocity network to predict the constant trajectory velocity.

Rectified flow: Rectified flow interpolates a clean image latent z0 and Gaussian noise latent z1 ∼ N(0, I) as zt = (1 − t)z0 + tz1 for t ∈ [0, 1].
Rectified flow: The velocity network vθ predicts z1 − z0 from noisy latent zt, timestep t, and text prompt c; a coarse clean-latent estimate can then be recovered from the predicted velocity.

4 Method

Delta-Adapter learns exemplar-based image editing from a single source-target pair by conditioning on a semantic transformation delta rather than directly exposing the target image. A Perceiver-based adapter injects this delta into a frozen editing backbone, while flow matching and semantic consistency losses supervise reconstruction and edit direction.

Single-pair formulation: Given exemplar pair {a, a′}, the model extracts the transformation a →a′ as a semantic delta and applies it to an unseen query image without textual guidance.The model takes (a, ∆a→a′) as input and reconstructs a′, allowing each single pair to supervise itself.
Training objective: The training objective combines flow matching reconstruction with semantic delta consistency to align the generated edit direction with the ground-truth transformation.The consistency term uses patch-weighted cosine distance, emphasizing regions whose ground-truth semantic deltas have larger magnitude; L = Lflow + λsdc Lsdc.
Semantic delta extraction: Patch-level SigLIP features are layer-normalized before differencing, then refined with a gated residual projection whose gate starts at zero.Layer normalization suppresses instance-level magnitude variation while preserving directional feature changes; the bounded learnable gate gradually incorporates residual corrections.
Delta-to-token adapter: The refined delta is converted into fixed-length edit tokens with a Perceiver resampler and per-token affine projections, preserving localized and relational edit information.Each latent token receives its own projection ei = Wi ri + bi, enabling distinct aspects of the edit to specialize.
Backbone injection: The edit tokens enter every DiT block through a decoupled cross-attention branch with learnable residual injection, while the backbone weights remain frozen.Only the delta-side projections and preceding projection layers are optimized.

5 Experiments

Experiments show that Delta-Adapter transfers exemplar edit semantics while preserving query-image content, outperforming baselines on seen and unseen tasks. It also supports continuous editing control and effective test-time adaptation for difficult unseen pairs.

Evaluation protocol: The evaluation uses LPIPS, CLIP-I, GPT-C, and GPT-A to assess perceptual similarity, semantic alignment, content consistency, and editing accuracy.GPT-C and GPT-A use a 5-point scale.
Qualitative evaluation: Delta-Adapter more faithfully captures exemplar edit semantics while better preserving the query image’s structure and identity in qualitative comparisons.
Continuous image editing: Delta-Adapter enables continuous image editing by adjusting the decoupled cross-attention injection strength λca.This control follows from injecting only the editing signal, unlike full-exemplar-pair methods such as RelationAdapter.
Quantitative evaluation: Excluding LoRWeB’s frequent near-identity outputs, Delta-Adapter achieves superior editing accuracy and content consistency over all baselines.LoRWeB attains the highest consistency scores but substantially lower editing accuracy because of frequent editing failures.
Test-time adaptation: 20 gradient steps of test-time adaptation substantially improve difficult unseen exemplars, changing coarse edits into outputs that faithfully reflect the specified transformation.

6 Conclusions and Limitations

Delta-Adapter enables exemplar-based image editing with single-pair supervision and no textual guidance by conditioning on a semantic delta instead of the edited image. The method eliminates paired exemplars during training and outperforms four strong baselines on seen and unseen editing tasks.

Conclusions: Delta-Adapter performs exemplar-based image editing under single-pair supervision without textual guidance.The framework conditions on a semantic delta rather than directly on the edited image.
Conclusions: Conditioning on a semantic delta eliminates the need for paired exemplars during training, enabling scalable data curation and test-time adaptation.
Conclusions: Delta-Adapter achieves superior performance over four strong baselines on both seen and unseen editing tasks.

A Implementation Details

Delta-Adapter is built on FLUX.2-klein-4B2 with SigLIP-2 image encoding and a Perceiver-resampler adapter, and is trained on approximately one million image pairs. Baselines use official checkpoints or retraining protocols, with text instructions generated for methods that require them.

Our method: Delta-Adapter uses FLUX.2-klein-4B2, SigLIP-2, and a Perceiver resampler with 128 learnable queries.Training uses AdamW with lr = 1 × 10−4, weight decay = 0.01, β1 = 0.9, and β2 = 0.99 in bfloat16 precision.
Our method: The full model trains for 100K steps on approximately one million image pairs, using per-GPU batch size 16 across 4× H200 GPUs.
Baselines: RelationAdapter, LoRWeB, and VisualCloze use official pre-trained checkpoints and default inference configurations, while Edit Transfer is retrained on the Relation dataset.Because these baselines require text alongside exemplar pairs, concise edit descriptions are generated for each evaluation case using GPT-5.4.

B Additional Qualitative Results

Additional qualitative results demonstrate Delta-Adapter’s performance on challenging edits and compare it with baseline methods across seen and unseen editing tasks.

Additional Qualitative Results: Figures 7–9 show challenging Delta-Adapter edits and qualitative comparisons with baselines on seen and unseen editing tasks.Figure 7 presents broader challenging results, while Figures 8 and 9 cover seen- and unseen-task comparisons, respectively.

C Additional Quantitative Results · D Comparison with Additional Baselines · E Failure Cases

Additional results show that Delta-Adapter outperforms baselines on an unseen validation set, while general-purpose multimodal models struggle with exemplar-based editing. Its main limitation is preserving fine-grained details, especially textual content and exemplar-consistent characters.

C Additional Quantitative Results: Delta-Adapter consistently outperforms all baselines in editing accuracy and content consistency on RelationAdapter’s unseen validation set.The evaluation uses CLIP-I and LPIPS consistency scores between generated and ground-truth images.
C Additional Quantitative Results: 4.077 GPT-A score for Delta-Adapter exceeds RelationAdapter’s 3.432, the best-performing baseline.These scores are reported on the unseen validation set.
D Comparison with Additional Baselines: Nano Banana 2 and GPT-Image-2 often fail to capture intended edits in exemplar-based image editing.The comparison covers rows 1–3 of Figure 10.
D Comparison with Additional Baselines: Nano Banana 2 and GPT-Image-2 tend to leak appearance cues from exemplar images into their outputs.This limitation is illustrated in rows 4–6 of Figure 10.
E Failure Cases: Delta-Adapter can struggle to preserve fine-grained visual details, particularly textual content, despite generalizing across many editing tasks.Figure 12 presents representative failure cases.
E Failure Cases: The model may generate characters inconsistent with those in the exemplar pair.The limitation is partly attributed to the pre-trained vision encoder’s inability to faithfully capture fine-grained visual details.

F Details of GPT-Based Evaluation Metrics

The evaluation uses GPT-5.4 to score edited results for editing accuracy and content consistency from four image inputs. It assesses transformation fidelity and preservation of non-edited query content.

GPT-5.4 serves as an automated evaluator for each edited result.
The evaluator reports two dimensions: editing accuracy (GPT-A) and content consistency (GPT-C).
Each evaluation receives the reference source a, reference target a′, query image b, and candidate edit ˆb′.
GPT judges whether the transformation a →a′ is faithfully applied to b and whether non-edited query content is preserved.

G Human Preference Evaluation · H Societal Impact · I Licenses for Pre-trained Models and Datasets

The paper evaluates Delta-Adapter through pairwise human preferences, discusses its applications and misuse risks, and documents the licenses of its pretrained models and datasets. Additional qualitative results, comparisons, and failure cases illustrate transfer quality, generalization, and limitations.

G Human Preference Evaluation: Human participants compare Delta-Adapter with each baseline by selecting which edit better transfers the exemplar transformation while preserving unedited query regions.The study collects 768 judgments.
H Societal Impact: Delta-Adapter enables exemplar-pair-guided editing without textual instructions, supporting content creation, artistic exploration, and visual prototyping.The visual-only formulation lowers the barrier to exemplar-based editing.
H Societal Impact: The method may be misused to create misleading imagery, motivating reliable detection of synthetic or edited images and responsible deployment practices.The passage identifies misuse mitigation as requiring detection and responsible deployment.
I Licenses for Pre-trained Models and Datasets: The implementation uses FLUX.2-klein-4B under the FLUX.2 Community License and SigLIP-2 under the Apache 2.0 License.These are the pretrained model licenses specified by the paper.
I Licenses for Pre-trained Models and Datasets: Training uses the publicly available Relation, Pico-Banana, and NHR-Edit image-pair datasets while adhering to their authors’ licenses.Evaluation images come from these datasets or from the cited external sources.
I Licenses for Pre-trained Models and Datasets: Additional qualitative results show Delta-Adapter inferring transformations from exemplar pairs and applying them to unseen query images.The figure presents additional editing results produced by the method.
I Licenses for Pre-trained Models and Datasets: On seen and unseen editing tasks, qualitative comparisons report more faithful semantic transfer and better preservation or generalization than the listed baselines.Seen-task comparisons include RelationAdapter, LoRWeB, and Edit Transfer; unseen-task comparisons include RelationAdapter, LoRWeB, and VisualCloze.
I Licenses for Pre-trained Models and Datasets: Delta-Adapter struggles with precise text rendering, producing characters inconsistent with those in exemplars containing textual content.This is identified as a failure case of the method.