acceptodds
Desk rejected at ICLR 2027

CrossPair: Unlocking Generalist Visual Transformation Learning in Image Editors

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Language models switch tasks from demonstrations, but vision still often relies on separate systems for restoration, editing, and stylization. A before–after example offers a shared instruction only if the transferable change can be separated from the scene that demonstrates it. We introduce CrossPair, which adapts a pretrained multimodal image editor to transfer diverse transformations from one exemplar pair. Exemplar-before, exemplar-after, and query images take fixed roles in its native VLM and VAE pathways: the VLM extracts operation semantics, while the VAE retains fine appearance and structure. A training-only, change-aware DINOv3 objective aligns feature-change directions across content-distinct scenes, guiding the output to follow the demonstrated relation without an inference-time encoder. We construct Transform-360K by grouping filtered pairs from conditional representations, restoration, editing, virtual try-on, and stylization under shared transformations. On Relation252K-derived seen and held-out transformations, \method improves target agreement and edit-direction similarity over released systems. Ablations show that the semantic and visual exemplar routes are complementary and that DINO relation alignment further strengthens transfer; Transform-360K training also improves aggregate held-out-operation scores.

Primary area:
generative models
Keywords:
Diffusion model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_YgHhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CrossPair, a generalist image editing framework designed to transfer visual transformations from a single before–after exemplar pair to a content-distinct query image. The central objective is to infer the transformation demonstrated by the reference pair and apply it to a new image while preserving unrelated query content. Built upon Qwen-Image-Edit, CrossPair introduces role-ordered exemplar conditioning through the pretrained editor's native VLM and VAE pathways. The VLM pathway provides semantic interpretation of the demonstrated operation, while the VAE pathway preserves fine-grained visual appearance and structural information. The authors further introduce a training-only DINOv3 relation alignment objective. This objective aggregates patchwise feature displacements using change-aware spatial weights and aligns transformation directions between the reference and query scenes without requiring direct spatial correspondence across different scenes. Additionally, the paper constructs Transform-360K, consisting of 91,010 four-image training records spanning more than 8,000 fine-grained transformation relations across restoration, stylization, conditional representations, virtual try-on, and spatial editing. Experiments on seen and held-out transformations derived from Relation252K show improvements in target agreement, CLIP directional similarity, and GPT-5.5-based transformation faithfulness compared with the evaluated released systems.

Critical strengths
  1. CrossPair makes effective use of the pretrained image editor’s existing VLM and VAE conditioning pathways. By feeding the reference-before, reference-after, and query images into both pathways in a fixed order, the method combines semantic understanding of the demonstrated transformation with the preservation of fine-grained visual details. Moreover, DINOv3-based relation alignment is used only during training, eliminating the need for an additional relation encoder or computational overhead at inference time.

  2. The construction of Transform-360K is a meaningful contribution. It contains 91,010 training records covering more than 8,000 fine-grained transformation relations, with operation-specific pairing and quality filtering across diverse editing tasks.

Critical weaknesses
  1. While CrossPair substantially improves transformation faithfulness, its query-content preservation remains considerably weaker than several baselines. For example, its in-domain content-preservation Rank-1 rate is only 15.3%, compared with 47.8% for LoRWeB (Table 2). This raises concerns about whether the model effectively transfers demonstrated transformations while preserving query-specific content.

  2. Although the change-aware DINOv3 relation objective is presented as a central technical contribution, Table 3 shows only marginal improvements in CLIP-I (GT) (0.776 to 0.778) and LPIPS (0.402 to 0.400), while the more noticeable gain is in CLIP-Dir (0.350 to 0.369). Moreover, the method introduces a DINOv3 ViT-7B/16 encoder during training, requiring additional feature extraction and gradient computation. Without an analysis of the additional training cost, it is difficult to assess whether these improvements justify the computational overhead.

  3. As described in the supplementary material, some prompts contain explicit editing descriptions in addition to image-role instructions. These descriptions may already specify the intended transformations, making it difficult to determine whether CrossPair learns transformations from visual exemplars or primarily follows textual instructions using a strong pretrained image editor. Since visual transformation learning is central to the paper, the contribution of visual exemplars should be evaluated more explicitly.

Questions for authors
  1. Could the authors report the additional training time, GPU memory usage, and computational overhead introduced by the DINO relation loss? Given its marginal improvements in CLIP-I (GT) and LPIPS, a cost–performance analysis would help establish its practical value.

  2. Could the authors evaluate CrossPair using role-only prompts without explicit editing descriptions? Additionally, how would the output change if the reference pair were replaced with one demonstrating a different transformation while keeping the query image and text instruction unchanged?

  3. Could the authors clarify why the full-model OOD results differ between Tables 2 and 3, particularly the CLIP-Dir scores (0.312 vs. 0.369)? Do these experiments use the same evaluation samples, model checkpoints, prompts, and inference configurations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer YgHh (1/2): Content Preservation and DINO Supervision

Thank you for recognizing our dual-pathway design for combining semantic understanding with visual detail, as well as the contribution of Transform-360K. This comment addresses Weaknesses 1-2 and Question 1. Comment 2/2 addresses Weakness 3 and Questions 2-3.

1. Content preservation and overall editing quality (Weakness 1)

(1) Does a lower content-preservation score mean worse editing?

Content preservation needs to be considered in the context of the requested transformation. Our benchmark includes style transfer, segmentation, and conditional representation tasks, where the correct output can differ substantially from the source. Style transfer changes texture and appearance, while segmentation or edge-map generation removes much of the photographic appearance. Higher source similarity can therefore also indicate that the requested transformation has not been fully applied.

On the seen tasks in Table 2, CrossPair has lower CLIP-I_src than LoRWeB (0.572 versus 0.690), but higher CLIP-Dir (0.438 versus 0.309), higher CLIP-I_GT (0.871 versus 0.772), and lower LPIPS_GT (0.361 versus 0.479). The outputs change the source appearance while becoming closer to the target and better following the demonstrated transformation.

Figure 5 provides concrete examples. For poster-style transfer, LoRWeB and RelationAdapter retain more of the train photograph's appearance. For watercolor transfer, LoRWeB also remains closer to a photographic appearance. CrossPair more fully reproduces the styles demonstrated by the exemplars.

(2) Does the overall evaluation support CrossPair's editing quality?

Yes. In Table 2, CrossPair achieves GPT Rank-1 rates of 53.7% for overall quality and 59.6% for transformation faithfulness, leading all compared methods on the seen tasks. LoRWeB's corresponding rates are 17.3% and 15.1%. Content preservation should therefore be evaluated alongside transformation fidelity. The target metrics, directional metric, and overall GPT evaluation all support more faithful transformation transfer and better overall editing results.

2. DINO benefits and computational cost (Weakness 2 and Question 1)

(1) What does DINO contribute when the gains in target similarity and LPIPS are small?

DINO relation supervision focuses on whether the query undergoes the same change as the exemplar. In Table 3, with the dual-pathway configuration unchanged, CLIP-Dir improves from 0.350 to 0.369, with improvements in target similarity and LPIPS as well. The main benefit is in directional consistency, which matches the objective's purpose. The dual-pathway model already has editing capability; relation supervision further improves transformation transfer.

(2) Where does the additional computation occur? Is DINO needed at inference?

During training, DINO's parameters are frozen and its inputs are resized to 224 x 224. Among the DINO branches, only the predicted-image branch retains input gradients through the encoder (Appendix C.5). Only the generator's LoRA parameters are updated. The additional training computation comes from decoding the predicted image, extracting DINO features, and backpropagating through the prediction branch.

DINO is not needed at inference. Appendix D explicitly states that DINO and the auxiliary loss computation are removed after training. Processing new exemplars and queries requires neither loading DINO nor running its feature extraction or relation computation. The benefit of this supervision therefore comes without additional inference-time DINO computation.

AuthorsOpenReview ↗

Response to Reviewer YgHh (2/2): Visual Exemplars and Evaluation Protocol

This comment addresses Weakness 3 and Questions 2-3. Comment 1/2 discusses content preservation and DINO's benefits and computational cost.

3. Text conditioning and the contribution of visual exemplars (Weakness 3 and Question 2)

(1) Does visual input still help when the text already describes the edit?

Yes. Section 3.2 and Appendix C.4 explicitly describe our text-assisted exemplar-editing setting. In Table 3, VLM-only and VLM+VAE use the same concrete edit text; the latter additionally supplies exemplars through the VAE. CLIP-I_GT improves from 0.768 to 0.776, CLIP-Dir from 0.343 to 0.350, and LPIPS_GT from 0.406 to 0.402.

These results show that VAE exemplar input provides an additional benefit when text and VLM exemplar information are already available. The improvement does not come from a more detailed textual instruction.

(2) How do we view evaluation with role-only prompts?

Role-only prompts can further assess the model's ability to infer the transformation from the visual exemplars alone. Our focus is on using visual exemplars more effectively in a text-assisted setting, so removing the concrete edit description entirely is not a prerequisite for establishing the pathway benefit above. We plan to include role-only evaluation in subsequent experiments and report the results in the appendix.

(3) What about changing the exemplar while keeping the query and text fixed?

The Relation252K test set includes multiple cases for the same transformation, with five queries per group in the seen-task evaluation. Table 2 shows how these transformations transfer across different query content. This provides evidence of cross-content transfer; changing only the exemplar while holding the query and text fixed is a further test of exemplar-specific control.

4. Differences between Tables 2 and 3 (Question 3)

(1) Why are the full model's OOD CLIP-Dir scores 0.312 and 0.369?

The tables use different evaluation sets. The component ablation in Table 3 uses the original 300 released Relation252K unseen cases, allowing all variants to be compared on the same fixed benchmark. We subsequently added 150 cases to broaden the coverage of transformation types in the main evaluation, giving the 450-case test set in Table 2. The original 300 cases are the established benchmark, not a subset selected from the 450 cases based on model performance.

Thus, 0.369 and 0.312 are averages over different evaluation sets. Section 4.1 and Appendix B.6 (Table 7) already describe the original 300 cases and the additional 150. We will state the sample count and its relationship to Table 2 directly in the Table 3 caption.

(2) Are the checkpoint, prompts, and inference settings the same?

Yes. The comparison is between CrossPair (R252K) in Table 2 and the full model in Table 3. They use the same checkpoint, prompts, and inference settings. The difference is the test set, rather than the model or inference configuration. All variants within Table 3 are evaluated on the same test set.

Dear Reviewer YgHh:

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your concerns. If so, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

Reviewer_9bmdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CrossPair infers a transformation from a before–after exemplar pair and transfers it to a content-distinct query image. Built on Qwen-Image-Edit, it supplies the exemplar-before, exemplar-after, and query images through the native VLM and VAE pathways. A training-only DINOv3 objective aligns feature-change directions across scenes. The paper also introduces TRANSFORM-360K, containing 91,010 four-image records. Evaluation on seen and unseen transformations shows improved target agreement and edit-direction similarity.

Critical strengths
  1. A unified interface supports diverse visual transformations. The method covers stylization, restoration, conditional representations, and other edits, demonstrating the practical potential of exemplar pairs as a general task interface.

  2. Relation supervision avoids unjustified cross-scene patch matching. Changes are localized within each scene before their aggregated directions are compared. This suits content-distinct exemplars and queries without requiring a DINO encoder at inference.

  3. Method comparisons are separated from new-dataset experiments. Relation252K-derived and TRANSFORM-360K models are reported separately, alongside matched-backbone component ablations. This helps establish that some gains arise from the method rather than exclusively from additional data.

Critical weaknesses
  1. Textual conditioning confounds exemplar-pathway attribution. Appendix C.4 shows that some inputs contain explicit edit descriptions, whereas the VAE-only ablation uses a generic prompt. The comparison therefore changes both exemplar routing and textual information. Without no-exemplar or incorrect-exemplar controls, dependence on the demonstrated relation remains unclear.

  2. Semantic exclusion of OOD operations requires clarification. Query exclusion and non-self exemplar sampling do not establish operation-level exclusion across training sources. Different sources may contain semantically equivalent transformations, requiring a clearer audit to support unseen-operation generalization.

  3. The necessity of the DINO relation objective is not fully established. In Table 3, directional similarity improves from 0.350 to 0.369, while target similarity and LPIPS improve only slightly. Without repeated experiments or comparisons against simpler perceptual supervision and unweighted change aggregation, the benefit of the specific relation design remains uncertain.

  4. Transformation transfer and content preservation exhibit a substantial trade-off. In Table 2, CrossPair has lower content-preservation Rank-1 rates and source similarity than major baselines. Some transformations legitimately require substantial changes, but aggregate metrics do not distinguish necessary edits from unintended content alterations.

Questions for authors
  1. Do the pathway-ablation conclusions hold with identical edit descriptions? How much does performance decline when exemplars are removed or replaced with incorrect transformations?

  2. How are semantically equivalent OOD operations excluded across all training sources? Can operation-level results be provided?

  3. Does the DINO objective outperform simpler perceptual supervision or unweighted variants? Are repeated-run results available?

  4. Which operations account for the lower content-preservation rankings? Is there evidence that non-target regions remain unchanged?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9bmd (1/2): Exemplar Conditioning and OOD Evaluation

Thank you for recognizing our unified treatment of diverse visual transformations, the design of relation supervision, and the separate evaluation of method and dataset contributions. We address each weakness together with its corresponding question. This comment covers Weaknesses 1-2 and Questions 1-2; comment 2/2 covers Weaknesses 3-4 and Questions 3-4.

1. Exemplar pathways and text conditioning (Weakness 1 and Question 1)

(1) Does the visual pathway still help when the text is unchanged?

Yes. In Table 3, VLM-only and VLM+VAE use the same concrete edit text and matched optimization settings. The latter additionally supplies exemplars through the VAE. CLIP-I_GT improves from 0.768 to 0.776, CLIP-Dir from 0.343 to 0.350, and LPIPS_GT from 0.406 to 0.402. All three metrics improve without adding textual information. This comparison shows that VAE exemplar input provides a benefit beyond the information already available through the text and VLM pathway.

VAE-only serves a different purpose: it reports performance with exemplars supplied only through the VAE and a generic instruction, without a concrete textual edit description. This prompt is disclosed in Appendix C.4. We therefore use the text-matched VLM-only/VLM+VAE comparison to assess the added VAE pathway, rather than attributing the entire VAE-only/VLM-only gap to routing. As Section 4.3 explains, these pathway labels concern exemplar inputs; the query retains its standard VLM and VAE pathways in all variants.

(2) Are tests with removed or incorrect exemplars necessary?

Our focus is how to use visual exemplars more effectively in the text-assisted setting stated in Section 3.2. Removing exemplars turns the task into text-guided editing; replacing them with exemplars that contradict the text also tests how the model handles conflicting conditions. These are useful supplementary tests, but they are not required to establish the pathway gain above.

Table 3 directly tests that gain under identical text. Table 2 evaluates transfer across different exemplar-query combinations and seen and unseen transformations. For example, the seen-task evaluation includes five queries per transformation group, testing whether the demonstrated change transfers to different image content.

2. OOD definition and exclusion across sources (Weakness 2 and Question 2)

(1) Is OOD defined separately against each model's training set?

Yes. The OOD transformations in the Relation252K and Transform-360K experiments are absent from each model's corresponding full training mixture, including those in the 150 added test cases. The Transform-360K evaluation does not simply reuse an "unseen" label defined only against Relation252K. The exclusion is relative to the complete training mixture used for that checkpoint.

Section 4.1 already defines OOD through held-out transformation groups. Query exclusion and non-self sampling in Appendix B.6 are additional sampling requirements. They do not replace transformation-level exclusion.

(2) How are equivalent transformations across sources treated?

The criterion is the actual transformation. Semantically equivalent transformations do not qualify as unseen merely because their sources or labels differ. This applies across all sources in each model's training mixture.

Section 3.5 describes the transformation grouping. These groups are more specific than broad task families. For example, a model may have seen style transfer but not the particular test style, virtual try-on but not the test garment, or reference segmentation but not the test target object. OOD cases are sampled evenly across the major task types, with these specific transformation forms held out. We will make this distinction and its separate application to both training mixtures more explicit in the experimental setup.

(3) Can results be reported by operation?

Table 2 reports aggregate results, while Appendix E.2 describes the per-query scores and per-task means produced by the evaluation. We plan to add a more detailed breakdown by transformation to the appendix in a subsequent version, so that differences across task types are easier to inspect.

AuthorsOpenReview ↗

Response to Reviewer 9bmd (2/2): Relation Supervision and Content Preservation

This comment addresses Weaknesses 3-4 and Questions 3-4. Comment 1/2 covers exemplar conditioning and OOD exclusion.

3. Contribution of DINO relation supervision (Weakness 3 and Question 3)

(1) Why is the main improvement in CLIP-Dir?

The main gain is in directional consistency, which matches the objective's purpose. Table 3 shows that relation supervision further improves transformation transfer; the dual-pathway model already has editing capability. DINO is used only during training and is not needed at inference.

(2) What do relation supervision and change weighting contribute?

A direct perceptual loss compares the prediction with the target image. Our relation loss compares the change in the query scene with the change demonstrated by the exemplar. It therefore supervises whether the same transformation is applied across different image content.

As described in Section 3.4, changes are localized separately within each scene before their pooled feature directions are aligned. This avoids requiring patch correspondence between unrelated scenes. Change weighting emphasizes the edited regions when pooling feature changes, reducing the influence of unchanged regions. Table 3 reports the overall benefit of adding this relation objective to the same dual-pathway model.

4. Transformation fidelity and content preservation (Weakness 4 and Question 4)

(1) Does lower source similarity mean worse editing?

This depends on the task. Our benchmark includes style transfer, segmentation, and conditional representation tasks, where the desired output can differ substantially from the source. Style transfer changes texture and appearance, while segmentation or edge-map generation removes much of the original photographic appearance. High source similarity can therefore also indicate that the requested transformation has not been fully applied.

This is why we report source similarity alongside CLIP-Dir and target agreement. CLIP-Dir checks whether the query change follows the exemplar's change, while CLIP-I_GT and LPIPS_GT measure agreement with the target output.

On the seen tasks in Table 2, CrossPair has lower CLIP-I_src than LoRWeB (0.572 versus 0.690), but higher CLIP-Dir (0.438 versus 0.309), higher CLIP-I_GT (0.871 versus 0.772), and lower LPIPS_GT (0.361 versus 0.479). Thus, the outputs change the source appearance while becoming closer to the target and better following the demonstrated transformation. These measures should be read together when judging editing quality.

(2) Do the visual comparisons show insufficient editing?

Yes. In the second row of Figure 5, LoRWeB and RelationAdapter retain more of the original train photograph's appearance, whereas CrossPair more closely follows the exemplar's illustrated poster style. In the third row, LoRWeB still retains more photographic appearance, while CrossPair reproduces the exemplar's watercolor treatment more clearly. These examples show why preserving more source appearance does not always mean completing the requested edit more faithfully.

(3) How does GPT assess overall editing quality?

Content preservation matters, but it needs to be considered together with whether the requested transformation is completed. Section 4.2 therefore discusses preservation alongside transformation fidelity and overall quality.

On the seen tasks in Table 2, CrossPair achieves GPT Rank-1 rates of 53.7% for overall quality and 59.6% for transformation faithfulness, leading all compared methods in both dimensions. For comparison, RelationAdapter reaches 24.0% and 21.3%, respectively. These results agree with the directional and target metrics and the visual comparisons: CrossPair more faithfully applies the demonstrated transformation and achieves better overall editing results.

Dear Reviewer 9bmd:

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your concerns. If so, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

Reviewer_qC2kOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors developed a visual example image editing network by fine-tuning a flow matching-based image editing method (Qwen-Image-Edit) with the visual example as additional input. Addtionally they utilized DINOv3 for a training objective that aligns the change direction between the query-output pair and the reference pair. A large-scale in-context dataset was compiled from a variety of available datasets for training and evaluation. The authors compare their method to two other baselines under seen-edit and OOD-edit setups.

Critical strengths
  • The authors utilize a powerful image editor as a base, which results in a big improvement in performance.
  • The authors compile a large-scale in-context dataset from a variety of available datasets, which can be quite useful for future work.
Critical weaknesses
  • The paper is badly written and hard to follow. The language used and sentence structures make it hard to understand without substantial effort.

  • Critical Issues with the experiments:

    • Limited benchmark: The authors only compare with two other methods that utilize different backbones. While it was important to fix the training data, it is not quite clear where the difference in performance is coming from because of the big difference between the backbone of each method. The authors need to specify what contribution they are trying to quantify and design a benchmark that highlights that contribution as the main difference in the comparison.

    • OOD comparison is not thoroughly established:

      • The huge number of transformation groups (8k) used can result in classifying edits with small differences as different classes. This can result in some of the OOD holdout classes being relatively similar to the seen distribution. There is no information from the authors about the choices made to ensure a valuable difference between seen and OOD.
      • There is overlap between the Relation252K dataset and the additional tasks added to create the TRANSFORM-360K. This can result in a mix between seen and OOD classes when utilizing both datasets for evaluation or training.
    • Missing correct evaluation for low-level and perception tasks: The paper claims low-level vision and image perception tasks as part of the tasks that their model can do. The authors provide no specific evaluation or proper metrics to show that.

  • Limited novelty: The authors include the visual example as just additional input without novelty in conditional efficiency, which results in a large inference overhead that increases the input size by 4x. Even though the use of DINO for direction matching is valuable, the main idea of matching directions is already utilized in other editing methods, including text-based ones. Lastly, the dataset is larger-scale than previous ones; however, it is a compilation of already existing data without additional scaling.

Questions for authors
  • What is the additional complexity (FLOPs and inference time) from all additional tokens added by the visual reference inclusion? What resources are required for training your model?
  • What is the performance of the model (specialized metrics) on low-level (Deblur, SR, ...) and perception tasks (Depth, Segmentation) compared to a specialized ref and other compared methods?
  • How are transformation groups classified, and how did they reach 8k? What kind of groups are used for OOD evaluation, and how do they differ from the seen groups?
3 comments
AuthorsOpenReview ↗

Response to Reviewer qC2k (1/2): Critical Weaknesses

Thank you for recognizing the performance gains from a strong base editor and our dataset's potential value for future research. This comment addresses your Critical Weaknesses; comment 2/2 answers your questions.

Weakness 1: Clarity of the paper

The problem we study is to apply the change shown by a before-and-after example to a different query image. The model needs to understand both what should change and what that change should look like.

Secs. 3.2-3.4 follow this idea: the VLM helps interpret the operation, the VAE supplies visual details such as color, texture, and structure, and the relation loss encourages the query's change to follow the example's change. We will make this explanation clearer in the method overview.

Weakness 2: Comparisons, OOD definition, and task evaluation

(1) How do we separate component gains from backbone differences?

As explained in Sec. 4.1, Table 2 compares complete systems, including frozen Qwen, LoRWeB, and RelationAdapter. Table 3 tests components within the same backbone.

In Table 3, adding VAE example input while keeping the edit text unchanged raises CLIP-Dir from 0.343 to 0.350. Adding DINO with both pathways fixed raises it further to 0.369. Target similarity and LPIPS also improve. These comparisons support the component contributions without assigning all cross-backbone gains to our design. The main R252K-trained model also does not use Transform-360K, so its gains cannot be explained by the new corpus.

(2) What distinguishes the 8K groups and the OOD cases?

The groups represent specific transformations, not 8K unrelated task families. A particular style, try-on with a particular garment, and reference segmentation of a particular target object each define a transformation.

OOD cases are sampled evenly across task types. The specific test transformations are absent from the corresponding training mixture. For example, training can include style transfer but not the test style; virtual try-on but not the test garment; or reference segmentation but not the test target object. The task family may be familiar while the specific transformation is unseen.

(3) Does overlap between data sources mix seen and unseen operations?

Table 1 discloses that Transform-360K includes Relation252K training records. Shared sources do not mean shared test transformations. The OOD transformations are absent from each model's respective training mixture, including those in the 150 added cases. For Transform-360K, this applies to its full training set, not just the Relation252K portion.

(4) How are low-level vision and perception outputs evaluated?

Our evaluation includes LPIPS and GPT-based rankings alongside the CLIP metrics. LPIPS measures perceptual distance to the target. The GPT judge uses the example pair to assess transformation faithfulness, query-content preservation, and overall preference. These provide a common evaluation across different transformations.

Task-specific metrics answer more detailed questions: perceptual distance and visual rankings do not replace depth error or segmentation mIoU. We plan to add more detailed per-task results and suitable task-specific metrics in the appendix of a future extended version.

Weakness 3: Method and dataset contributions

(1) What is new beyond adding reference images or using a directional loss?

Our focus is how to use the example pair to transfer a change across different contents. The before image, after image, and query take fixed roles in both native pathways. The VLM supplies operation semantics, while the VAE retains fine visual details. The relation loss then trains the generator to apply the demonstrated change to the query.

Eqs. (7)-(9) first locate changes separately within each scene, then pool feature changes and align their directions. This avoids assuming that patches in unrelated scenes correspond. The query target supplies detached localization weights; the predicted change is aligned with the example's change rather than directly matched to target features. Table 3 tests the gains from the added VAE route and relation loss. Sec. 2.2 already acknowledges that single-example transfer is established prior work.

(2) Is Transform-360K only a collection of existing files?

Appendix B.3 describes additional pair construction. For style data without paired source images, we generate photorealistic counterparts. We generate conditional representations and construct forward, reverse, and selected cross-condition mappings. The try-on pipeline uses generated paired targets. We then filter the pairs and form training records with different contents under the same transformation.

The work therefore includes pair generation, transformation grouping, quality checks, and cross-content pairing. Appendix B.2 also states the counting unit: 91,010 four-image records contain 364,040 record-level image slots, not 360K independent images.

AuthorsOpenReview ↗

Response to Reviewer qC2k (2/2): Questions for Authors

This comment replies to your three Questions for Authors. Our replies to the Critical Weaknesses are in comment 1/2.

Question 1: Training resources and inference time

(1) What resources were used for training?

Training used 8 PPUs. Appendix C gives the other training settings, including bfloat16, ZeRO-2, gradient checkpointing, and rank-128 generator LoRA. Only the generator's LoRA parameters are updated; the original model parameters remain frozen.

(2) How much inference time does CrossPair require?

Under the same comparison settings, CrossPair takes approximately twice the inference time of [REFERENCE MODEL]. This overhead comes from processing the example conditions. DINO is not part of inference: the entire branch is removed after training, so there is no DINO model to load or DINO feature extraction to run.

Question 2: Low-level vision and perception performance

(1) What do the current results measure?

The current evaluation uses LPIPS to compare outputs with their targets, and GPT-based rankings to assess the demonstrated transformation and query-content preservation. Together with the CLIP metrics, these give a shared evaluation across tasks. Table 2 reports comparisons with the listed baselines under these measures.

(2) Will there be more detailed task-specific comparisons?

The paper focuses on using one example-based interface for different visual transformations. We plan to add more detailed per-task results, suitable task-specific metrics, and comparisons with relevant specialist methods in the appendix of a future extended version. These would give a fuller picture of restoration and perception performance beyond the current aggregate results.

Question 3: Transformation groups and OOD sampling

(1) How are transformations grouped?

We first organize the data into broad task types, then distinguish specific transformations within each type. Examples include:

  • Style transfer: a particular style defines a transformation.
  • Virtual try-on: try-on with a particular garment defines a transformation.
  • Reference segmentation: segmentation of a specified target object defines a transformation.
  • Conditional representations: a directed mapping, such as RGB-to-depth, defines a transformation; the reverse direction is different.
  • Restoration: groups follow the degradation task.
  • Spatial editing: groups follow the specified viewpoint, rotation, or distance change.

Sec. 3.5 and Appendix B describe these grouping rules and the data construction.

(2) Why are there more than 8K relations?

Each task type can contain many styles, garments, target objects, or geometric changes. The number of broad task types is therefore much smaller than the number of specific transformations. The 8K figure counts these fine-grained relations, not 8K high-level tasks. Appendix B.2 also distinguishes relation counts from source-local serialized labels, which should not be counted directly as independent relations.

(3) How are OOD cases sampled, and how do they differ from training cases?

OOD cases are sampled evenly across the task types. The key difference is that the specific transformation is absent from the corresponding model's training mixture. For example:

  • The model has trained on style transfer, but not the particular test style.
  • It has trained on virtual try-on, but not the particular garment used at test time.
  • It has trained on reference segmentation, but not the particular target object used at test time.

This tests transfer to unseen transformations within the task types. It does not require every OOD case to belong to a completely new task family.

Dear Reviewer qC2k:

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your concerns. If so, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

Reviewer_8D7ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to infer a visual transformation from one before–after exemplar pair and apply it to a query image depicting different content. CrossPair feeds the exemplar-before, exemplar-after, and query images in a fixed role order through both the native VLM and VAE conditioning pathways of Qwen-Image-Edit-2511, while training only generator LoRA weights. A training-time DINOv3 relation objective estimates changed regions within each scene, computes differences between normalized patch features, and forms a change-weighted pooled vector that is then normalized before exemplar/query directions are aligned. The DINO branch is removed at inference.

The authors also construct Transform-360K, comprising 91,010 four-image records, or 364,040 record-level image slots, drawn from restoration, conditional representations, virtual try-on, style, and spatial transformations. The name “360K” does not denote the number of independent images. Experiments separately evaluate a primary model trained on a Relation252K-derived mixture and a Transform-360K-trained model on seen and designated unseen transformations. Target agreement, CLIP directional similarity, and single-judge rankings broadly support the system's editing capability. However, the independent contribution of the visual exemplars and operation-level generalization after expanding the training corpus need more direct verification.

Critical strengths
  1. The training-time relation objective has a clear motivation and a coherent implementation. The method uses separate within-scene change weights and aligns pooled relation directions across content-distinct scenes, avoiding a direct cross-scene patch-correspondence assumption (§3.3–3.4, Equations (5)–(11), Figure 2). DINO is removed at inference, so this supervision does not require an additional auxiliary encoder during generation. The contribution is the specific conditioning/supervision recipe; single-exemplar transfer itself is established prior work.

  2. The reported results include both substantial system-level gains and useful within-backbone component evidence. Without the new corpus, CrossPair improves over RelationAdapter on seen tasks from 0.777 to 0.871 in CLIP-I_GT, from 0.355 to 0.438 in CLIP-Dir, and from 0.429 to 0.361 in LPIPS_GT. The main OOD changes are 0.722→0.744, 0.253→0.312, and 0.476→0.440. Table 3 additionally reports matched optimization: adding the VAE exemplar route to VLM exemplar conditioning changes CLIP-I_GT from 0.768 to 0.776, CLIP-Dir from 0.343 to 0.350, and LPIPS_GT from 0.406 to 0.402; adding DINO then gives 0.778, 0.369, and 0.400. These are positive component results, although their uncertainty is not reported. The authors also disclose weaker content-preservation rankings and the complete-pipeline nature of the external-baseline comparison.

Critical weaknesses
  1. Exemplar-specific control beyond the concrete edit text remains insufficiently characterized, and the VAE-only versus VLM-only comparison is text-confounded. The paper explicitly permits and discloses concrete editing text; its use is not itself a flaw or a violation of the stated setting. However, Appendix C.4 specifies a generic fallback prompt for the VAE-only ablation, whereas the other relevant records retain concrete edit semantics. The VAE-only/VLM-only gap therefore cannot be attributed solely to exemplar routing. These labels concern the exemplar routes: all variants retain the query's standard VAE/VLM routes (§4.3). This concern does not invalidate the cleaner VLM-only→dual-pathway or dual-pathway→DINO comparisons. A compact experiment holding a coarse text instruction and query fixed while changing the exemplar's visual details would more directly establish what the exemplar contributes beyond text. Table 10 also configures the frozen Qwen baseline with exemplar input only in the VLM route; a full-triplet frozen dual-pathway control would separate interface and fine-tuning effects. The authors already label Table 2 a complete-system comparison, so this asymmetry limits causal attribution rather than invalidating that comparison.

  2. The OOD scope relative to Transform-360K needs more explicit clarification. Section 4.1 explicitly defines unseen transformations as held-out transformation groups, and Appendix B.6 describes query exclusion and non-self sampling. I accept this as the reported protocol; the submission does not establish an overlap or leakage incident. The remaining ambiguity is whether the same operation-level exclusion is enforced separately with respect to both the Relation252K-derived and Transform-360K training mixtures, including equivalent operation keys across their sources and the 150 added cases. Clarifying this would distinguish generalization to operations still unseen after corpus expansion from improved performance on a benchmark originally designated unseen relative to Relation252K. A clear exclusion description and compact operation inventory would address this concern; public release of all manifests or images is not a prerequisite.

Questions for authors
  1. Exemplar contribution and ablation prompts: Could you clarify the pathway comparison when equivalent concrete edit semantics are supplied to the VAE-only and VLM-containing variants? On a representative existing test subset, holding the query and a coarse text instruction fixed while varying valid exemplars with different visual details would help demonstrate exemplar-specific control. A frozen Qwen control receiving the full triplet through both pathways would also help separate interface and fine-tuning effects. Text-only or role-only tests could be useful supplementary diagnostics if the interface supports them, but a drop under an unfamiliar input format alone should not be treated as conclusive evidence of exemplar dependence.

  2. Unseen relative to which training mixture: Does the held-out-transformation statement in §4.1 apply independently to both checkpoints? Please describe how equivalent operations across source datasets are excluded, including the 150 added cases. If some operations are covered by T360K, separate those from operations still unseen relative to T360K. A concise inventory/exclusion explanation is sufficient; this is not a demand for a new dataset or a full public data release.

  3. Reproducibility and incremental effects: Please specify the evaluation membership/count and checkpoint used in Table 3, and the total optimization steps or equivalent training exposure for the reported configurations. The text identifies a Relation252K unseen-task OOD split, but does not explicitly give Table 3's size or its exact relationship to the 450-query main panel. Per-query or per-transformation paired differences with suitable uncertainty would help interpret the reported DINO changes (+0.002 CLIP-I_GT, +0.019 CLIP-Dir, −0.002 LPIPS_GT). If available, a uniform-weight aggregation control would further isolate change weighting; the current numbers alone do not establish statistical significance or insignificance.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8D7Z (1/2): Critical Weaknesses

This comment responds to the two Critical Weaknesses in your review. A separate comment (2/2) addresses the three Questions for Authors.

Thank you for the careful review and for distinguishing the system-level results from the within-backbone evidence.

Weakness 1: Exemplar contribution, edit text, and the frozen Qwen baseline

(1) What the pathway comparisons establish.

The VAE-only row evaluates a specific configuration: the exemplar is supplied through visual latents without a concrete textual edit description, while the query retains its standard VAE/VLM routes. It reports how this configuration performs; Appendix C.4 explicitly documents the generic prompt.

The evidence that VAE exemplar input adds value comes from a different comparison: VLM-only versus VLM+VAE. Both retain the same concrete edit text and VLM exemplar information. Adding the VAE exemplar route improves CLIP-I_GT from 0.768 to 0.776, CLIP-Dir from 0.343 to 0.350, and LPIPS_GT from 0.406 to 0.402. Thus, even after the model receives the edit description and semantic exemplar representation, the additional visual input improves the result. These gains cannot be explained by providing more detailed text.

With both pathways fixed, adding DINO further gives 0.778 / 0.369 / 0.400. These two comparisons support the additional value of visual conditioning and relation supervision separately. As you note, they are not affected by the VAE-only prompt difference. Our component conclusions therefore do not require attributing the entire VAE-only/VLM-only gap to routing. We will make the prompt settings and these comparison pairs clearer in Table 3.

(2) Exemplar control and the scope of the existing evaluation.

The central task is to apply a demonstrated transformation to different query content. Relation252K evaluates multiple queries for the same transformation: the seen-task panel has 219 groups with five queries each (Sec. 4.1). Table 2 reports target agreement and directional similarity across these cases, providing evidence of transfer across different query contents.

Your proposed fixed-query, fixed-text exemplar swap asks a more specific question: whether visual details absent from the text follow the selected exemplar. Our cross-query results do not isolate that intervention. The existing evidence supports the value of visual exemplar input under the disclosed text-assisted setting; it should not be read as a claim that text and exemplar contributions have been fully separated.

(3) Why we retain Qwen's original input interface.

In the Qwen-Image-Edit implementation we use, multiple images enter the VLM, while the VAE receives only the image being edited. We preserve that design for the frozen baseline: exemplar-before, exemplar-after, and query enter the VLM, and only the query enters the VAE. This keeps the baseline faithful to the base system rather than incorporating CrossPair's dual-pathway exemplar configuration before comparison.

Accordingly, Sec. 4.1 explicitly describes Table 2 as a complete-system comparison, and Table 10 lists the interfaces. Table 3 then examines component effects within the same backbone. A frozen full-triplet dual-pathway control would further separate interface and fine-tuning effects; we do not attribute the entire Table 2 gain over Qwen to fine-tuning alone.

Weakness 2: OOD scope relative to each training mixture

(1) Operation exclusion applies independently to both checkpoints.

Yes: the transformations evaluated as OOD are absent from the corresponding checkpoint's training mixture, independently for the Relation252K-derived and Transform-360K models. This includes the transformations in the 150 added cases. The T360K designation is not simply inherited from Relation252K's unseen split.

As stated in Sec. 4.1, the held-out unit is a transformation group, not merely a new image. Query exclusion and non-self sampling in Appendix B.6 are additional sample-construction requirements; they do not replace operation-level exclusion.

(2) What counts as the same operation across sources.

The criterion is the actual transformation, not its dataset name or source-local label. Sec. 3.5 defines relations by style or garment identity, directed modality mapping, restoration task, or geometric change. A different label for the same operation does not make it unseen. A broad task family may occur in training while a particular transformation within it is held out; we do not equate unseen transformation groups with entirely unseen task families.

Table 1 lists the Relation252K training records included in Transform-360K. Shared provenance therefore does not imply inclusion of the held-out operations. The OOD statement applies against each checkpoint's complete training mixture, including the additional sources in T360K. We will state this scope directly in Sec. 4.1 and Appendix B.6 so that readers need not infer it from separate descriptions.

AuthorsOpenReview ↗

Response to Reviewer 8D7Z (2/2): Questions For Authors

This comment replies to your three Questions for Authors. Our replies to the Critical Weaknesses are in comment 1/2.

Question 1: Example inputs and ablation prompts

(1) What happens when the edit text is the same?

The same-text comparison in Table 3 is VLM-only versus VLM+VAE. Adding the VAE example input improves CLIP-I_GT / CLIP-Dir / LPIPS_GT from 0.768 / 0.343 / 0.406 to 0.776 / 0.350 / 0.402. The text is unchanged, so these gains show the benefit of the additional visual input.

The VAE-only row answers a different question: how well the model performs with the example pair supplied only through the VAE and a generic prompt. It is not a same-text comparison with VLM-only.

(2) What evidence relates to changing the example?

Relation252K includes multiple queries for each transformation, with five queries per seen group. Table 2 shows transfer across these different query images. This tests cross-content transfer, but does not measure what happens when only the example changes while the query and text stay fixed.

(3) Why not give frozen Qwen the full triplet through both pathways?

We kept the multi-image design of the Qwen-Image-Edit implementation we use: multiple images enter the VLM, while only the query enters the VAE. This lets Table 2 compare the original base system with CrossPair. Table 3 separately tests the effect of adding VAE example input. A frozen dual-pathway control would further separate the effect of the input setup from fine-tuning.

(4) What about text-only or role-only tests?

Text-only means keeping the query and edit instruction but removing the example pair. Role-only means keeping all three images, with an instruction such as "Apply the change from image 1 to image 2 to image 3," without naming the edit.

Our setting allows text to assist example-based editing (Sec. 3.2). We ask whether visual input still helps when the text already describes the edit, and Table 3 supports this. Role-only testing would further measure how well the model infers the edit from the images alone.

Question 2: Unseen relative to which training mixture?

(1) Does the statement apply to both checkpoints independently?

Yes. The OOD transformations are absent from the respective training sets of both the Relation252K-derived and Transform-360K models. The same rule applies to the 150 added cases. A transformation is not treated as unseen by Transform-360K merely because it belongs to Relation252K's unseen split.

(2) How are equivalent operations across sources treated?

The actual transformation determines whether it is seen. Changing its dataset name or label does not make it a new operation. Sec. 3.5 describes the operation definitions and directions. We will make it explicit in Sec. 4.1 that the exclusion applies separately to each model's full training set, including all sources used by Transform-360K.

Question 3: Table 3 and the DINO improvement

(1) Which cases are used in Table 3?

Table 3 uses the original 300 released Relation252K unseen cases. All variants use this same test set. We later added 150 cases to broaden the main evaluation, giving the 450 cases in Table 2. The 300 cases are the original benchmark, not cases selected from the larger set based on performance.

Sec. 4.1 and Appendix B.6 describe the original and added cases. We will also put the count and its relation to Table 2 directly in the Table 3 caption.

(2) Are the checkpoint, prompts, and inference settings the same?

Yes. CrossPair (R252K) in Table 2 and the full model in Table 3 use the same checkpoint, prompts, and inference settings. The scores differ because they summarize different test sets, rather than two evaluations on the same set.

(3) What does DINO improve?

DINO trains the model to make the query's change follow the example's change. With both input pathways fixed, CLIP-Dir rises from 0.350 to 0.369. CLIP-I_GT also rises by 0.002, and LPIPS_GT falls by 0.002. The main improvement is in the change direction, which is what the loss is designed to encourage (Sec. 3.4).

DINO is used only during training. At inference, we remove the whole branch: there is no DINO model to load and no DINO feature extraction to run.

Dear Reviewer 8D7Z:

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your concerns. If so, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.