Edit Spillover as a Probe: Do Image Editing Models Implicitly Understand World Relations?

Guandong Li; Zhaobin Chu

cs.CVMar 18, 2026

Edit Spillover as a Probe: Do Image Editing Models Implicitly Understand World Relations?

Authors: Guandong Li, Zhaobin Chu

Organizations: iFLYTEK

Abstract

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.

Figures & tables

Category	Count	Description
App Screenshots	54	Mobile app interfaces with structured layout
Normal Docs	23	Documents, posters with mixed text/images
Others	13	Diverse scenarios
Real Scenes	57	Photos with signboards, road signs
Receipts	53	Invoices with high-density text

Table 1: EditSpilloverBench dataset composition (200 images)

Model	Architecture	Rationale
Nano Banana	Production text editor	Spillover discovery source
Qwen-2511	Dual-stream DiT	High quality representative
FireRed	Qwen-based	Architecture comparison
SeedReam 4.5	Large pretrained	High spillover case
WanX 2.6	Tongyi series	Medium spillover case

Table 2: Evaluated models

Model	Imgs	Spill% $\downarrow$	SSIM $\uparrow$	Regions	Spat.	Sem.	Mix.	Rand.	WUS	Sem.Cnt	Sem.Den. $\uparrow$
nano_banana	200	11.46	0.869	89.0	17.5	31.2	12.9	38.3	1.69	5561	27.8
seedream45	198	11.12	0.864	240.8	21.1	26.7	7.0	45.1	1.21	12730	64.3
firered	200	7.55	0.901	79.6	12.9	36.4	15.0	35.7	2.62	5796	29.0
wanx26	192	6.92	0.894	104.5	15.8	34.6	14.6	34.9	2.06	6950	36.2
qwen_2511	198	3.49	0.942	40.2	9.4	40.5	11.2	38.9	3.87	3222	16.3

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

\times

from near to far distance).

Distance	nano_banana	seedream45	firered	wanx26	qwen_2511
0–0.5 $\times$	100	100	100	100	100
0.5–1.0 $\times$	38.4	65.4	70.9	94.8	216.7
1.0–1.5 $\times$	19.5	41.3	45.6	62.4	102.8
1.5–2.0 $\times$	12.8	21.8	24.5	38.4	64.3
2.0–3.0 $\times$	5.8	15.0	16.7	21.6	50.5
3.0–5.0 $\times$	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 $\times$	53.2	58.5
0.5–1.0 $\times$	42.4	56.4
1.0–1.5 $\times$	39.0	51.1
1.5–2.0 $\times$	38.7	50.9
2.0–3.0 $\times$	46.5	52.5
3.0–5.0 $\times$	41.7	48.6

Table 5: Semantic spillover proportion (%) by distance

$\beta$	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

\beta

(

\alpha=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

\beta=0.80

$\alpha$	nano	seed	fire	wanx	qwen
1.0	3.42	2.45	5.36	4.36	8.03
1.5	1.69	1.21	2.62	2.06	3.87
2.0	1.05	0.78	1.60	1.33	2.35

Table 8: WUS values under different

\alpha

(

\beta=0.80

)

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 315 papers

Image-Space Rule Discovery
Aug 1, 2026Misora Sugiyama, Toya Oyama, Hirokatsu KataokaImage Editing Evaluation Text-Guided Image Editing
The University of Tokyo · National Institute of Advanced Industrial Science and Technology (AIST) · Visual Geometry Group, University of Oxford
Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing
Apr 16, 2026Yixuan Ding, Wei Huang, Ruijie Quan +2Image Editing Evaluation Text-Guided Image Editing
Zhejiang University · The University of Hong Kong
Semantic Granularity Navigation in Image Editing
May 20, 2026Liangsi Lu, Minzhe Guo, Xuhang Chen +1Diffusion Model Sampling Diffusion Timestep Scheduling
Guangdong University of Technology, Guangzhou, China · Huizhou University, Huizhou, China.

Aug 1, 2026J/K move · Enter open · S save

Image-Space Rule Discovery

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

Image Editing Evaluation Text-Guided Image Editing

Open paper AskPDF arXiv

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.

Abstract

Image Editing

Explore similar work

Image-Space Rule Discovery

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

Semantic Granularity Navigation in Image Editing