Instruction-following image editing models are expected to modify only the specified region while keeping the rest of the image unchanged. However, in practice, we observe a pervasive phenomenon -- edit spillover: models alter semantically related but unspecified content outside the edit region. This raises a fundamental question -- does spillover reflect genuine implicit world understanding, or is it merely attention leakage? We propose EditSpilloverProbe, a systematic framework that repurposes edit spillover as a natural probe for world knowledge in image editing models. We introduce a spillover taxonomy (spatial, semantic, mixed, random), an automated detection-and-classification pipeline, and a benchmark dataset constructed from real-world Chinese text editing tasks, EditSpilloverBench. Systematic evaluation of 5 representative editing models reveals three core findings: (1) spillover rates vary dramatically across architectures, from 3.49% to 11.46%, with a 3.3x ratio; (2) absolute semantic spillover quantity reveals models' world understanding capability -- nano_banana produces the most semantic spillover (27.8 per image), while qwen_2511 has the most precise editing control but lower semantic spillover (16.3 per image), revealing a trade-off between editing control and world understanding; (3) spatial decay analysis shows spillover area density decays exponentially with distance, but the proportion of semantically relevant spillover remains constant (40%-58%), providing direct evidence that semantic spillover reflects genuine world understanding rather than spatial diffusion.
Table 3: Spillover analysis results. Spill%: spillover rate (%), Spat./Sem./Mix./Rand.: spillover type proportions (%), Sem.Cnt: total semantic regions, Sem.Den.: semantic density (regions/image).
Figure 1: Cross-model comparison radar chart. Models exhibit different trade-offs between spillover control (SSIM, low Spill%) and world understanding (WUS, Semantic Density).
Figure 2: Spillover area density decay with distance. All models exhibit exponential decay, with nano_banana showing the steepest decline (200 × from near to far distance).
Distance
nano_banana
seedream45
firered
wanx26
qwen_2511
0–0.5 ×
100
100
100
100
100
0.5–1.0 ×
38.4
65.4
70.9
94.8
216.7
1.0–1.5 ×
19.5
41.3
45.6
62.4
102.8
1.5–2.0 ×
12.8
21.8
24.5
38.4
64.3
2.0–3.0 ×
5.8
15.0
16.7
21.6
50.5
3.0–5.0 ×
2.1
5.7
4.8
6.6
21.3
Table 4: Spillover area density decay with distance (normalized to near distance = 100%). Distance is measured in multiples of edit box diagonal.
Figure 3: Semantic spillover proportion remains constant across distances (39%–58%), providing direct evidence that semantic spillover is not driven by spatial proximity.
Distance
nano_banana
qwen_2511
0–0.5 ×
53.2
58.5
0.5–1.0 ×
42.4
56.4
1.0–1.5 ×
39.0
51.1
1.5–2.0 ×
38.7
50.9
2.0–3.0 ×
46.5
52.5
3.0–5.0 ×
41.7
48.6
Table 5: Semantic spillover proportion (%) by distance
β
nano
seed
fire
wanx
qwen
0.70
8.38
9.13
17.09
15.72
25.67
0.75
3.81
3.35
7.16
5.71
10.79
0.80
1.69
1.21
2.62
2.06
3.87
0.85
0.66
0.44
0.88
0.71
1.49
0.90
0.21
0.13
0.37
0.25
0.66
Table 6: WUS values under different β ( α=1.5 )
Model
Spat.
Sem.
Mix.
Rand.
nano_banana
17.5
31.2
12.9
38.3
seedream45
21.1
26.7
7.0
45.1
firered
12.9
36.4
15.0
35.7
wanx26
15.8
34.6
14.6
34.9
qwen_2511
9.4
40.5
11.2
38.9
Table 7: Classification distribution (%) at β=0.80
Can image editors solve worksheets while preserving unrelated content? WISP (Worksheet Image-Space Problem Solving) defines 11 tasks and eight information conditions. On 2,200 shared items per editor, Nano Banana Pro achieves a 48.7% strict automatic proxy pass rate (Auto-Strict), versus 0.0-13.4% for four other editors. Analysis of 13,200 outputs, including a separate diagnostic, distinguishes target, preservation, and format failures. Auto-Strict agrees with human-majority judgments on 95.5% of 440 outputs (inter-rater Fleiss' κ=0.894). Excluding the two no-edit controls, Nano Banana Pro achieves 45.1% Auto-Strict, versus at most 3.2% for the other editors. Further probes show high symbol-tracking proxy scores but only 49% exact Wason selection. These results measure problem solving through edits, without establishing open-ended rule induction or transfer.
Misora Sugiyama, Toya Oyama, Hirokatsu Kataoka
The University of Tokyo · National Institute of Advanced Industrial Science and Technology (AIST) · Visual Geometry Group, University of Oxford
Diffusion-based image editing has achieved strong visual fidelity under natural language instructions, yet most existing systems still operate at the level of surface instruction following, without reasoning about the implicit contextual constraints embedded in real user requests. This often leads to visually plausible but logically inconsistent edits. In this work, we introduce RE-Edit, a benchmark for REasoning-aware image Editing that evaluates image editing systems across five complementary reasoning dimensions: physical, environmental, cultural, causal, and referential. RE-Edit comprises 1,000 carefully curated samples, each designed such that visual plausibility alone is insufficient and correct editing requires satisfying implicit logical constraints. To support fine-grained analysis, we establish dimension-aligned evaluation criteria and conduct a comprehensive study of ten open-source and two commercial image editing models. Our results show that even advanced systems frequently struggle with implicit multi-dimensional reasoning despite producing high-quality visuals. We further present a lightweight reasoning-guided post-edit baseline as an initial exploration, illustrating how inserting explicit reasoning can help mitigate such failures in a model-agnostic manner.
Despite the generative capabilities of diffusion and flow models, real-image editing remains constrained by a persistent trade-off between semantic editability and structural fidelity. We trace a primary cause of this limitation to the implicit coupling of edit progress with model scale in existing paradigms. Under this coupling, stronger edits typically require visiting noisier states, which spends computation on destabilizing layout before the semantic change is well localized. We introduce NaviEdit, a training-free inference-time controller that decouples edit progress from model scale traversal through a strict self-consistency contract. NaviEdit operates at the rollout level and leaves the underlying pretrained model unchanged. It treats scale as a control input and reallocates a fixed step budget toward semantically responsive intermediate scales instead of destructive high-noise regimes. Experiments show positive average gains across compatible editors and flow backbones, supporting decoupling as a portable inference-time control principle.
Liangsi Lu, Minzhe Guo, Xuhang Chen +1
Guangdong University of Technology, Guangzhou, China · Huizhou University, Huizhou, China.