Text-to-image (TTI) models increasingly generate high-quality images from natural-language prompts, yet figurative language exposes a failure: a vehicle that should guide the depiction of a tenor may instead be rendered as a visible object. We call this failure Figurative Vehicle Intrusion: the intruding content is textually licensed, but it is assigned the wrong visual role, showing that visual presence is not always faithfulness and that presence-oriented evaluation can miss such errors. To study it systematically, we introduce Vehicle Intrusion and Semantic Tenor Assessment (VISTA), a multilingual benchmark of figurative prompts organized by Figurative Form and Mapping Mechanism. We further propose V-Score, a diagnostic question-answering metric that evaluates role-aware figurative faithfulness in generated images. Evaluations on recent high-performing TTI models show that vehicle intrusion persists across languages and figurative categories. As a lightweight mitigation, we introduce VISTA-Guard, which partially reduces vehicle intrusion and suggests a practical path toward more figuratively faithful TTI generation. All resources will be released publicly.
Figures & tables
Figure 1: Figurative vehicles literalized by models.
Figure 3: Overview of VISTA Benchmark Construction and Evaluation.
Annotation
zh
en
ja
ar
Total
Simile
130
220
190
120
660
Metaphor
120
110
230
120
580
Substitutive Metaphor
120
90
20
160
390
Idiom
70
70
60
100
300
Semantic Shift
180
320
270
200
970
Perceptual Resemblance
100
60
110
100
370
Table 1: VISTA sample counts by language and annotation view.
Figure 4: Base performance on VISTA. (a) Sample-level V-Score distributions for each model. (b–d) Mean V-Score with standard errors across Languages, Figurative Forms, and Mapping Mechanisms.
Model
Tenor Preservation
Vehicle Avoidance
Scene Coherence
Figurative Meaning
3
2
1
0
3
2
1
0
3
2
1
0
3
2
1
0
Nano Banana Pro
72.5
23.3
4.1
0.0
31.6
11.4
41.5
15.5
75.1
19.7
4.7
0.5
79.3
18.1
2.1
0.5
GPT-Image-2
76.7
19.2
4.1
0.0
33.2
5.7
40.4
20.7
71.5
22.8
4.1
1.6
86.0
11.4
2.1
0.5
DALL-E-2
69.3
26.6
4.2
0.0
32.3
6.8
37.5
23.4
69.8
22.9
7.3
0.0
81.8
15.6
2.1
0.5
DALL-E-3
72.4
22.4
5.2
0.0
35.4
7.3
37.0
20.3
72.9
18.8
7.3
1.0
84.4
13.5
2.1
0.0
Grok Imagine
77.2
18.7
3.6
0.5
35.8
10.9
35.8
17.6
69.9
22.8
5.7
1.6
77.7
17.6
4.1
0.5
Table 2: Option-score distributions for each model on Base generation. Values are percentages of valid samples assigned to scores 3–0 for each diagnostic question.
Model
TIFA-like
V-Score
Score
Rank
Score
Rank
Nano Banana Pro
93.50
4
77.17
2
GPT-Image-2
98.50
1
76.25
4
DALL-E-2
94.50
2
73.83
5
DALL-E-3
94.50
2
76.50
3
Grok Imagine
92.00
5
78.50
1
Table 3: TIFA-like and V-Score evaluation results.
Figure 5: Sample-level V-Score distributions for Base and VISTA-Guard across five generators.
Model
Human
Base
Generic Anti-Hall
VISTA-Guard
Score
Δ Score
Δ Q2
Score
Δ Score
Δ Q2
Nano Banana Pro
78.2
79.1
79.5
+0.42
-0.13
87.1
+7.94
+1.08
GPT-Image-2
77.3
78.9
80.2
+1.33
+0.21
88.7
+9.86
+1.18
DALL-E-2
78.2
77.4
79.7
+2.37
-0.09
86.3
+8.96
+1.25
DALL-E-3
78.8
78.9
79.4
+0.54
-0.27
86.5
+7.57
+1.13
Grok Imagine
79.9
79.3
81.2
+1.90
+0.06
84.5
+5.18
+0.76
Table 4: V-Score under four prompting strategies. Δ Score and Δ Q2 denote changes in the overall score and vehicle-avoidance score relative to Base, respectively.
Figure 6: Two representative cases comparing Base and VISTA-Guard. Prompts are translated into English.
Figure 7: Human–MLLM agreement on 30 blind images.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Language
Family
Writing system
Segmentation
Direction
Chinese
Sinitic
Logographic
No whitespace word segmentation
left-to-right
English
Germanic
Alphabetic
Whitespace-delimited words
left-to-right
Japanese
Japonic
Mixed scripts
No whitespace word segmentation
left-to-right
Arabic
Semitic
Abjad
Morphologically rich
right-to-left
Appendix
Table 12
Label
Core distinction
Decision boundary
Simile
Explicit comparison
Tenor and vehicle are linked by a marker equivalent to “like” or “as.”
Metaphor
Implicit attribution
Vehicle characterizes an explicit or recoverable tenor without an overt comparison marker.
Substitutive Metaphor
Referential replacement
Vehicle expression stands in for, or partially replaces, the tenor, which must be recovered from context ( Zhang and Gelernter, 2015 ) .
Idiom
Lexicalized form
Conventionalized multiword expression conveys a non-compositional meaning ( Sag et al., 2002 ) .
Appendix
Table 13
Label
Core distinction
Decision boundary
Semantic Shift
Conceptual transfer
Interpretation depends on abstraction, concretization, or related conceptual movement rather than perceptual similarity.
Perceptual Resemblance
Shared perceptual property
Mapping is licensed by visible or audible similarity, such as shape or sound ( Perniss et al., 2010 ; Ćwiek et al., 2021 ) .
Cross-modal Mapping
Sensory transfer
A property associated with one modality characterizes an experience in another, as in synesthetic expressions ( Duffy and Simner, 2010 ) .
Idiomaticity
Conventional meaning
Interpretation relies on lexicalized, non-compositional knowledge rather than a productive conceptual or perceptual mapping ( Sag et al., 2002 ) .
Appendix
Table 14
Annotation
zh
en
ja
ar
Total
Simile
13
22
19
12
66
Metaphor
12
11
23
12
58
Substitutive Metaphor
12
9
2
16
39
Idiom
7
7
6
10
30
Semantic Shift
18
32
27
20
97
Perceptual Resemblance
10
6
11
10
37
Appendix
Table 5: Composition of the 193-sample evaluation subset.
Model
Q1+Q2
Q3+Q4
Equal
Default
Nano Banana Pro
71.24
90.93
81.09
79.12
GPT-Image-2
70.64
91.19
80.92
78.86
DALL-E-2
68.84
90.19
79.51
77.38
DALL-E-3
70.83
90.97
80.90
78.89
Grok Imagine
72.88
88.95
80.92
79.31
Appendix
Table 6: V-Score sensitivity to different aggregation schemes on Base generation.
Model
Main Evaluator
Qwen Evaluator
Score
Rank
Score
Rank
Nano Banana Pro
77.17
2
75.72
2
GPT-Image-2
76.25
4
74.32
3
DALL-E-2
73.83
5
71.15
5
DALL-E-3
76.50
3
74.10
4
Grok Imagine
78.50
1
76.35
1
Appendix
Table 7: Cross-evaluator robustness of V-Score on the 40-prompt subset.
Figure 8: Image-level relationship between TIFA-like score and V-Score on 200 images. Points are slightly jittered for visibility.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Similes provide a compact and expressive way to describe visual characteristics in text prompts. Recent text-to-image models (t2i models) can produce visually compelling outputs from simile prompts, yet even frontier models frequently misinterpret the metaphorical vehicle and confuse it with the object. These systematic failures reveal a gap between figurative language and object-level visual grounding in t2i models. To investigate this issue, we propose a scalable evaluation framework for simile understanding. Our framework includes (1) a controlled simile dataset in which metaphorical vehicles are drawn from a predefined set of object-detectable categories and combined with diverse templates, (2) automatic grounding metrics based on YOLO (You Only Look Once) detection, and (3) text encoder layer analysis using Diffusion Lens to track how metaphorical vehicles emerge during generation. Experiments across architecturally diverse t2i models reveal consistent literalization failure patterns. We further discuss potential mitigation strategies for improving simile grounding in t2i models.
Text-to-image models are typically reported on average-case prompts, which understates the gap between systems on compositionally demanding requests involving precise object counts, multi-object attribute binding, legible embedded text, and explicit spatial constraints. We evaluate four production text-to-image systems: Hunyuan 3.0, Gemini 3 Pro Image ("Nano Banana Pro"), Black Forest Labs FLUX.2, and Ideogram 3.0. The evaluation uses the 48 hardest prompts drawn from the DataSeeds.AI Sample Dataset (DSD), selected through an automated complexity-scoring pass over the full corpus. Every generated image is graded using an independent-judge rubric. GPT-5.4-Pro authors an atomic, weighted, mutually exclusive and collectively exhaustive (MECE) evaluation rubric, while Gemini 3.1 Pro Preview independently determines whether each criterion is satisfied. Gemini 3 Pro Image ranks first with a score of 84.8/100, narrowly ahead of FLUX.2 at 82.3/100. Ideogram 3.0 and Hunyuan 3.0 score 65.7/100 and 63.3/100, respectively. Failure analysis shows that the leading systems primarily lose points through object miscounting and geometric artifacts, whereas the trailing systems more frequently produce garbled text. Ideogram 3.0 also frequently omits requested elements. Full per-sample rubrics, scores, and failure annotations are available from the authors upon request.
Recent text-to-image (T2I) models can render highly legible and well-structured text within images, enabling applications including document generation and slide generation. However, it remains unclear whether such systems faithfully preserve reasoning ability when complex solutions must be expressed directly through rendered text, or whether they merely imitate surface-level patterns. We investigate this question by evaluating reasoning fidelity in visual text generation, where models must express complete reasoning processes as images. Our evaluation includes long text rendering, factual knowledge probing, context understanding, and multi-step reasoning. Across these settings, we find that current T2I models frequently produce semantic errors, logical inconsistencies, and incorrect intermediate steps, even when the rendered text appears visually clear. These failures contrast with the strong reasoning performance of text-only models on the same tasks. Our findings reveal a substantial gap between visual text generation and procedural reasoning, motivating more reliable visual text reasoning.