CrossPair: Unlocking Generalist Visual Transformation Learning in Image Editors

• A before–after pair reveals both the transformation and invariant content, but the transferable rule must be inferred from one demonstration and applied to another scene.Source: §1 · p. 1
• The reference-before, reference-after, and query images take fixed roles in native VLM and VAE pathways that combine operation semantics with fine visual detail.The VLM interprets the operation, while the VAE retains colors, textures, shapes, and local structures.Source: §1 · pp. 1–2
• A training-only, change-aware DINOv3 objective aligns relation directions across scenes without requiring an inference-time encoder.It localizes informative changes independently in each scene before aligning their feature-change directions.Source: §1 · pp. 1–2
• TRANSFORM-360K contains 91,010 four-image records spanning more than 8,000 relations.Source: §1 · p. 2
• CROSSPAIR improves target agreement and edit-direction transfer on seen and held-out tasks.Source: §1 · p. 2

2.1 Diffusion-Based Image Editing

• Text-guided editing and spatial controls specify tasks through language or auxiliary maps, while reference-conditioned methods transfer appearance or object content.Source: §2.1 · p. 2

2.2 Visual In-Context Editing

• One-pair transfer is established; CROSSPAIR distinguishes itself with fixed triplet conditioning in native pathways and change-weighted relation regularization without an inference-time auxiliary encoder.Source: §2.2 · p. 2, §2.1 · p. 2

3.1 Rectified Flow

• Rectified flow predicts target-latent velocity from a noisy latent, timestep, and condition, yielding a differentiable one-step estimate of the clean latent.Source: §3.1 · pp. 2–3

3.2 One-Pair Inference, Cross-Pair Supervision

• At inference, one reference pair and a query image guide prediction, while training uses an aligned query target as cross-pair supervision.The paper distinguishes one-pair prompting from single-pair supervision, which uses the reference pair alone.Source: §3.2 · p. 3

3.3 Role-Ordered Exemplar Conditioning

• The ordered reference-before, reference-after, and query triplet enters both VLM and VAE pathways, combining operation semantics with fine-grained appearance evidence.The VLM abstracts the operation; VAE tokens retain color, texture, and local structure.Source: §3.3 · p. 3, §3.2 · p. 3

3.4 Change-Aware DINO Relation Alignment

• Change-aware DINOv3 alignment independently localizes informative changes, aggregates within-scene feature displacements, then aligns relation directions across scenes without patch correspondence or direct target-feature matching.Query-target-derived localization weights are detached, and the frozen DINOv3 branch is used during training only.Source: §3.4 · pp. 3–4

3.5 TRANSFORM-360K

• Transform-360K normalizes examples by their finest verifiable operation and direction, then validates visible edits, content consistency, and source-specific requirements.Validation also screens readability, resolution, duplicates, and changes that are trivial or unbounded.Source: §3.5 · p. 4
• 67,271 of 72,273 captioned-source records and 7,552 of 12,450 MegaStyle candidates survive their respective construction audits.Source: §3.5 · p. 4
• 91,010 four-image training records pair distinct content within operation keys across more than 8,000 fine-grained relations, with relation frequency and source families balanced.The manifest combines an exemplar pair with an aligned query pair.Source: §3.5 · p. 5

4.1 Evaluation Setup

• The protocol tests 219 seen transformation groups and 450 OOD queries, where unseen means held-out transformation groups rather than new images from seen groups.The comparison uses released systems with their native inference procedures, making it a complete-system comparison rather than a mechanism-matched backbone study.Source: §4.1 · p. 5

4.2 Results

• CrossPair has the highest in-domain target agreement and edit-direction similarity, while LoRWeB leads on source similarity and content-preservation preference.The TRANSFORM-360K checkpoint improves aggregate OOD target and directional scores; GPT-5.5 rankings favor CrossPair on transformation faithfulness.Source: §4.2 · pp. 6–7, Table 2 · p. 6

4.3 Ablation Study

• VAE-only conditioning has the highest source similarity, VLM-only scores better on target and direction, and combining routes raises target agreement beyond either alone.With both routes fixed, DINO relation supervision gives the strongest directional score and lowest target LPIPS in Table 3.Source: §4.3 · p. 7, Table 3 · p. 7

5 Conclusion

• CrossPair and Transform-360K support a shared demonstration-based interface for diverse visual transformations.The method uses training-only DINOv3 relation alignment and role-ordered VLM/VAE conditioning; the corpus checkpoint broadens the range of held-out operations addressed.Source: §5 · p. 7
• Exemplar-guided editing can transfer harmful changes, and synthetic or web-derived examples may carry visual biases.The paper recommends evaluating both transformation transfer and retained query content alongside deployment.Source: §5 · p. 7

A Evaluation Scope

• CrossPair’s Relation252K-derived model is the core method comparison, while the Transform-360K checkpoint assesses the combined effect of a larger, differently composed training corpus.Released systems are compared as complete pipelines; matched ablations examine CrossPair’s components.Source: A Evaluation Scope · p. 10
• CLIP and LPIPS are query-level micro averages across task families, while GPT-5.5 provides one ranking per query.Source similarity and content ranking complement target agreement in Table 2.Source: A Evaluation Scope · p. 10
• The aggregate measures characterize the sampled benchmark; operation-specific accuracy and pixel-level preservation require task-specific annotations.Source: A Evaluation Scope · p. 10

B.1 Training-unit schema

• Each record contains a demonstration pair, a query pair, and an instruction; the pairs share a construction group but are content-disjoint, and each query uses one non-self demonstration.Relation252K prompts can include a role instruction and a concrete edit description.Source: §B.1 · p. 10

B.2 Frozen manifest statistics

• The frozen manifest contains 91,010 records with four image slots each, or 364,040 record-level slots rather than unique files.The scan finds 248,910 unique path strings and 15,428 prompt strings; its serialized labels are not the paper-level count of over 8K fine-grained relations.Source: §B.2 · p. 10

B.3 Source-specific pair construction

• Source construction includes style-transfer pairs, conditional and segmentation representations, try-on records grouped by garment, spatial transformations, low-level vision groups, and imported Relation252K pairs.The style streams include MegaStyle, OmniStyle, Civitai-derived, and other records; the try-on manifest has 13,527 records over 3,936 garment labels.Source: §B.3 · pp. 10–11

B.4 Quality control

• MegaStyle’s semantic audit retained 7,552 of 12,450 pairs, while the captioned-record preflight retained 67,271 of 72,273 after decoding or geometry failures were removed.The checks also remove small segmentation masks and entries on a manually curated style bad-list.Source: §B.4 · p. 11

B.5 Grouping and record formation

• The assembler groups pairs by operation key and selects a non-self peer as demonstration; the frozen manifest uses K = 1 and yields 364,040 image slots from 91,010 records.Source-specific builders enforce distinct-content or scene constraints; low-level records also exclude references with the same internal group ID.Source: §B.5 · p. 11

B.6 Evaluation sets

• The seen-task builder reserves the last five eligible images per task, while the OOD protocol combines 300 released unseen cases with 150 additional cases from diverse transformation families.Tasks with at most 19 images are excluded, and held-out queries are excluded from their corresponding training groups.Source: §B.6 · p. 11

C.1 Architecture and trainable parameters

• CrossPair starts from Qwen-Image-Edit-2511 and optimizes only rank-128 LoRA adapters on the image generator, with the VLM, VAE, base generator, and DINOv3 frozen.Source: §C.1 · p. 12

C.2 Optimization configuration

• Training draws a scheduler index uniformly, adds Gaussian noise with scale 1, and uses scheduler-weighted MSE against the native flow-matching target.The checked configuration uses cfg_scale=1 and ConstantLR without warmup or decay; the K-bucket sampler gives each rank one record per step when K = 1.Source: §C.2 · p. 12

C.3 Resolution and image preprocessing

• The frozen training path preserves aspect ratio, rounds dimensions down to multiples of 16, and uses bilinear longest-side resizing without random crop, flip, or color jitter.The VLM references, VAE references, and DINO inputs use separate resolution targets.Source: §C.3 · p. 12

C.4 Role prompt and VLM serialization

• For K = 1, the VLM serializes the demonstration before, demonstration after, and query source in that order, followed by the record-specific prompt.The VAE-only ablation uses a fallback prompt and supplies only the query source to its VLM branch.Source: §C.4 · p. 12

C.5 DINOv3 implementation details

• The frozen DINOv3 ViT-7B/16 encoder retains 196 patch tokens, and the implementation uses detached, nonnegative change scores without cross-scene patch correspondence.In source_gt mode, the query map localizes the edit but does not replace the differentiable source-to-prediction direction.Source: §C.5 · p. 12

D Inference Details

• At inference, the exemplar-before, exemplar-after, and query images pass through both native pathways, while DINOv3, the target image, and auxiliary loss code are removed.The VLM serialization places three image placeholders before the record-specific prompt.Source: D Inference Details · pp. 12–13

E.1 Baseline provenance and inference protocol

• The corpus-controlled comparison uses Relation252K-trained CrossPair and RelationAdapter checkpoints; the base editor has no CrossPair LoRA weights, and LoRWeB includes additional training data.Methods retain their public generation interfaces, while metric inputs share a 512-pixel maximum long edge after output extraction.Source: §E.1 · p. 13

E.2 Deterministic image metrics

• CLIP-IGT measures cosine similarity between generated and target image features, while CLIP-Isrc compares the generation with the query source.The evaluation uses a specified Hugging Face CLIP checkpoint and L2-normalizes image features.Source: §E.2 · p. 13
• LPIPS uses an AlexNet backbone; generated images are resized to target dimensions with Lanczos interpolation before evaluation.The evaluation command defaults to a 512-pixel maximum long edge, a value explicitly set in the reported launchers.Source: §E.2 · p. 14
• The main table uses a query-level micro average, giving each valid query one value, rather than a transformation-level macro average.No confidence interval or significance claim is made for deterministic metrics.Source: §E.2 · p. 14

E.3 Blinded GPT-5.5 ranking

• GPT-5.5 ranks anonymized candidate outputs against the reference pair and query source, without receiving target images or method names.Candidate order is shuffled once per query; each query receives one request with four in-domain or five OOD candidates.Source: §E.3 · p. 14
• Every accepted ranking must include each anonymized candidate exactly once without ties, and the completed intersections contain 1,091 in-domain and 428 OOD queries, each judged once.Images are standardized before transmission, and failed calls are retried up to three times.Source: §E.3 · p. 14
• The evaluator reports Rank 1–K percentages and mean rank for each criterion, with query-level bootstrap confidence intervals for Rank-1.It also exports Rank-1 frequency by display position to audit residual position bias.Source: §E.3 · p. 14

F DINO Relation Diagnostics and Ablations

• Training logs track relation-loss activity and feature similarities alongside change-weight and gradient diagnostics; these are diagnostics, not image-quality metrics.The logged values include prediction–demonstration, target–demonstration, and prediction–target cosine similarities, plus the weighted auxiliary-to-primary output-gradient norm ratio and cosine.Source: F DINO Relation Diagnostics and… · p. 14