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.
Figures & tables
Figure 1: Single-pair visual transformation with CrossPair . A before–after exemplar specifies an operation for a content-distinct query.
Figure 2: Overview of CrossPair . The reference pair and query jointly enter the native VLM and VAE conditioning routes to guide flow prediction. During training only, the query target supplies detached change-localization weights and a frozen DINOv3 encoder compares reference and predicted-query relation directions; the VAE/VLM encoders are frozen and only generator LoRA weights are trained.
Manifest segment
Records
Share
Manifest segment
Records
Share
Low-level vision
20,049
22.03%
Style transfer sources
15,039
16.53%
Relation252K
17,600
19.34%
Virtual try-on
13,527
14.86%
Conditional representations
11,645
12.80%
Reference and segmentation
10,443
11.47%
Spatial transformations
2,144
2.36%
Other paired sources
563
0.62%
Total: 91,010 records (100%)
Table 1: Source composition of the frozen Transform-360K training manifest.
Figure 3: Transform-360K construction: source-specific pairing, quality filtering, and cross-pair formation within operation keys.
Deterministic metrics
GPT-5.5 Rank-1 rate (%)
Method
CLIP-I GT↑
CLIP-I src↑
CLIP-Dir ↑
LPIPS GT↓
Overall
Trans.
Content
(a) In-domain: 219 seen transformation groups
LoRWeB ( Manor et al., 2026 )
0.772
0.690
0.309
0.479
17.3
15.1
47.8
RelationAdapter ( Gong et al., 2025 )
0.777
0.602
0.355
0.429
24.0
21.3
22.3
Qwen Edit ( Wu et al., 2025 )
0.715
0.599
0.302
0.614
5.0
4.0
14.7
CrossPair (R252K)
0.871
0.572
0.438
0.361
53.7
59.6
15.3
Table 2: In-domain and OOD metrics with single-run GPT-5.5 Rank-1 rates. R252K and T360K denote the two training mixtures.
Figure 4: We compare CrossPair with LoRWeB, RelationAdapter, and Qwen-Image-Edit-2511 on in-domain Relation252K.
Figure 5: OOD comparison of Relation252K- and Transform-360K-trained CrossPair with released baselines.
OOD component ablation on Relation252K
Configuration
CLIP-I GT↑
CLIP-I src↑
CLIP-Dir ↑
LPIPS GT↓
VAE exemplar
0.711
0.685
0.279
0.426
VLM exemplar
0.768
0.618
0.343
0.406
VAE+VLM exemplar
0.776
0.620
0.350
0.402
+ DINO relation
0.778
0.622
0.369
0.400
Table 3: OOD component ablation; the final two rows isolate the DINO relation loss.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Manifest segment
Records
Share
Styles
Slots
Paths
Low-level vision
20,049
22.03%
20,049
80,196
80,196
Relation252K
17,600
19.34%
1 (220 task )
70,400
16,994
Virtual try-on
13,527
14.86%
3,936
54,108
51,436
Conditional representations
11,645
12.80%
196
46,580
14,402
Segmentation/reference perception
10,443
11.47%
9,723
41,772
18,387
MegaStyle
7,552
8.30%
7,552
30,208
30,208
Appendix
Table 5: Composition of the frozen Transform-360K manifest. Counts in the fourth column are serialized labels, not the paper-level relation count; segment-local labels can overlap.
Family
Manifest groups before final mixture assembly
Denoising
SIDD ( Abdelhamed et al., 2018 ) (80)
Deblurring
GoPro ( Nah et al., 2017 ) (63), RealBlur ( Rim et al., 2020 ) (234), DPDD ( Abuolaim & Brown, 2020 ) (250)
Resolution / moiré
Flickr2K ( Lim et al., 2017 ) (1,325), UHDM ( Yu et al., 2022 ) (1,260)
Weather removal
haze (746), rain (392), snow (750), raindrop (505) ( Li et al., 2019 ; Li et al., 2022 ; Liu et al., 2018 ; Qian et al., 2018 )
Illumination / color
LOLv2 ( Wang et al., 2022 ) (894), six Rendered-WB subsets ( Afifi & Brown, 2020 ) (853 each), UIEB ( Li et al., 2020a ) (346)
Reflection / shadow
SIR2 ( Wan et al., 2017 ) (72), SRD ( Qu et al., 2017 ) (1,543)
Appendix
Table 6: Low-level group counts in the checked construction manifest.
Item
Setting
Optimizer
AdamW, weight decay 0.01, β=(0.9,0.999) , ϵ=10−8
Learning rate / schedule
10−4 / ConstantLR
LoRA rank / target
128 / modules listed above
Per-device batch / accumulation
1 record / 1 step
Precision / distributed optimizer
bfloat16 / DeepSpeed ZeRO-2, no CPU offload
Gradient checkpointing
enabled
Appendix
Table 8: Training configuration.
Method
Visual input interface
Steps
Guidance
Seed / output handling
Qwen Edit 2511
one pair in the VLM route; query source remains in both native routes
40
4.0
42+i ; native output size
LoRWeB
one pair and query assembled as a 2×2 grid; lower-right output crop
repo. default
2.5
50; crop restored to source size
RelationAdapter
two conditioning images plus query-source VAE input
24
3.5
1000; native output
CrossPair
one pair and query through both VLM and VAE routes
40
4.0
42+i ; native output size
Appendix
Table 10: Checked baseline inference interfaces. Each method retains its released sampler; all outputs share the same metric preprocessing and evaluation resolution.