In visual storytelling, a believable character must not only look consistent across scenes but also behave in ways that reflect their personality. Current text-to-image (T2I) models preserve visual appearance effectively, yet consistently fail to maintain personality-relevant behavior---a failure we term trait collapse. We show that conventional metrics (CLIP Score, DINO similarity) are blind to this failure, and that abstract personality descriptors and scene-level emotion labels do not produce consistent improvements over neutral prompts across models. To systematically diagnose trait collapse, we introduce ViStoryBench-Personality, a story-level benchmark combining an expert-validated Visual Cues Lexicon (Fleiss' κ=0.757) with a six-dimensional metric suite covering personality alignment, cross-scene stability, and behavioral coherence. Evaluating seven T2I systems, we confirm a persistent semantic gap between abstract trait labels and visual evidence, and find that visual identity consistency does not imply personality consistency. To bridge this gap, we propose Trait-Behavior-Visual (TBV), a training-free prompting framework that decomposes personality traits into renderable visual constraints. TBV improves Personality Alignment Score by up to +22.1%, with larger gains in models with stronger language-conditioning capacity.
Figures & tables
Figure 1: Trait collapse vs. trait expression. Given the same narrative scene, baseline prompts (C1, left) produce generic or emotionally dominated behavior, while TBV prompts (C4, right) translate the OCEAN profile into concrete visual constraints, maintaining personality-consistent expression. Bottom: the TBV framework decomposes abstract personality traits into behavioral tendencies and then into renderable visual cues via an expert-validated lexicon.
Figure 2: Overview of the Trait-Behavior-Visual (TBV) framework. Phase T anchors the character’s OCEAN personality profile; Phase B derives context-sensitive behavioral tendencies via a self-correcting loop; Phase V maps behaviors into concrete visual constraints through the expert-validated Visual Cues Lexicon.
CLIP Score ↑
DINO Sim. ↑
Model
C1
C4
Δ
C1
C4
Δ
SDXL
.274
.262
−.012
.443
.475
+.032
FLUX.1-dev
.256
.187
−.069
.501
.504
+.003
SD3.5-Large
.260
.252
−.008
.467
.497
+.030
StoryDiff.
.259
.256
−.003
.378
.392
+.014
Qwen-2512
.263
.259
−.004
.644
.631
−.013
Table 1: CLIP Score and DINO similarity under C1 (baseline) and C4 (TBV). CLIP Score decreases under TBV prompting, likely due to prompt-length effects on CLIP-based text encoders; DINO similarity changes only marginally, with a mean signed change of +0.014 and a maximum absolute change of 0.032 . Neither metric captures the personality alignment improvements detected by our benchmark metrics (Table 2 ).
Model / Setting
PAS ↑
PCS ↑
RS ↑
CCS ↑
PDS ↓
SCS ↑
Open-source models: baseline
SDXL
0.446
0.686
0.429
0.357
+0.001
0.585
StoryDiffusion
0.480
0.763
0.467
0.499
+0.000
0.587
FLUX.1-dev
0.464
0.731
0.469
0.385
+0.003
0.614
SD3.5-Large
0.466
0.743
0.493
0.478
+0.003
0.605
Qwen-Image-2512
0.518
0.790
0.546
0.456
+0.001
0.665
Table 2: Personality consistency evaluation across seven T2I models. TBV yields up to +22.1% relative PAS improvement on Qwen-Image-2512, with larger gains observed in models equipped with stronger text encoders.
Model
C1
C4
Δ
p
Qwen-Image-2512
0.534
0.607
+0.073
.021
FLUX.1-dev
0.540
0.599
+0.059
.040
SDXL
0.426
0.481
+0.065
n.s.
Table 3: TBV gains (C4 vs. C1 PAS) under an independent judge (Gemini 2.5 Flash). Results are based on 80 matched C1–C4 story-shot pairs per model. Gains replicate in direction across all three models, with the largest gain preserved on Qwen-Image-2512. p from paired one-sided Wilcoxon signed-rank tests; n.s. = not significant.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Gemini Δ
p
GPT-4o Δ
n
Qwen-Image-2512
+0.073
.021
+0.075
80
FLUX.1-dev
+0.059
.040
+0.091
80
SDXL
+0.065
.133
+0.063
80
Appendix
Table 4: PAS gains (C4 − C1) on the same sampled shots under the independent judge (Gemini 2.5 Flash) and the primary judge (GPT-4o). p from paired one-sided Wilcoxon signed-rank tests on the Gemini scores. Directions agree across all models.
Pole
Hit Rate (%)
# Cues
Lift
O+
68.8
16
6.9 ×
O −
20.3
15
2.0 ×
C+
78.3
15
7.8 ×
C −
74.2
16
7.4 ×
E+
67.2
16
6.7 ×
E −
68.8
14
6.9 ×
Appendix
Table 5: Per-pole exact-match accuracy ( n=18 ). All poles exceed the 10% chance baseline. Low poles of O, A, and N show lower accuracy, reflecting genuine psychological overlap between conceptually adjacent dimensions (e.g., O − frequently mapped to C+).
Signal Type
Hit Rate (%)
# Cues
Facial expression
48.4
26
Body posture
59.0
25
Clothing
63.2
34
Color palette
61.4
15
Scene environment
54.5
50
Appendix
Table 6: Dimension mapping accuracy by signal type. Clothing cues show the highest discriminability; facial expression cues are most ambiguous.
α
Kendall’s τ
p
0.2
0.905
.003
0.3
1.000
< .001
0.4
1.000
—
0.5
1.000
< .001
0.6
1.000
< .001
0.7
1.000
< .001
Appendix
Table 7: Kendall’s τ between model rankings at each α and the paper’s α=0.4 . Rankings are fully stable across the central range.
Statistic
Value
Total stories
80
Total sequential shots
1,317
Unique characters
80
Mean shots per story
16.5
Evaluation shots
213
Conditions per shot
4 (C1–C4)
Appendix
Table 8: Overview of the ViStoryBench-Personality dataset.
Category
Count
%
Neutral
44
20.7
Low-arousal positive
41
19.2
Low-arousal negative
42
19.7
High-arousal positive
40
18.8
High-arousal negative
46
21.6
Visual difficulty
Appendix
Table 9: Distribution of affective scene categories and visual difficulty across 213 evaluation shots.
Figure 3: Big Five personality distribution across unique characters. Red = low pole (1–3), grey = neutral (4), blue = high pole (5–7).
Cond.
Prompt content
C1
Alice wishes she could shrink to fit through the small door. Daytime, inside the hall, disappointed and hopeful atmosphere. A curious girl with golden hair and blue eyes.
C2
C1 + [Personality: High Openness, Low Conscientiousness, High Extraversion, Low Neuroticism]
C3
C2 + feeling Disappointment
C4
C1 + animated facial micro-movements ∣ wide open exploratory gaze ∣ forward-leaning curious posture ∣ relaxed open hand gestures ∣ slight head tilt suggesting interest
Appendix
Table 10: Example prompt construction across four abstraction levels for Alice (O=7, C=3, E=5, A=4, N=3). C1–C3 test non-visual textual information, while C4 tests concrete visual grounding.
Model
Steps
CFG
Resolution
Arch.
SDXL
28
7.5
1024 × 1024
UNet+CLIP
FLUX.1-dev
28
3.5
1024 × 1024
DiT+T5
SD3.5-Large
28
4.5
1024 × 1024
MMDiT+3Enc
Qwen-2512
50
4.0
1328 × 1328
MMDiT+LLM
StoryDiff.
50
5.0
1024 × 1024
SDXL+CSA
GPT Image 2
—
—
1254 × 1254
Closed
Appendix
Table 11: Generation parameters for the seven T2I models.
Model
Device
Time/image
SDXL
RTX 4090
∼ 8 s
FLUX.1-dev
RTX 4090
∼ 12 s
SD3.5-Large
RTX 4090
∼ 15 s
Qwen-2512
A100
∼ 25 s
StoryDiffusion
A100
∼ 20 s
GPT Image 2
API
∼ 15 s
Appendix
Table 12: Approximate per-image generation time.
Component
Est. Cost (USD)
GPT Image 2 generation
∼ $240
Nano Banana generation
∼ $240
VLM evaluation (GPT-4o)
∼ $180
RTX 4090 GPU hours
∼ 200 h
A100 GPU hours
∼ 20 h
Total API cost
∼ $660
Appendix
Table 13: Estimated cost breakdown for the full evaluation pipeline. GPU hours reported separately as institutional compute.
Metric
Value
Spearman ρ
0.611
Pearson r
0.613
MAE (7-point scale)
0.871
Within ± 1 point
81.3%
Polarity agreement
86.6%
Appendix
Table 14: VLM-human agreement on personality evaluation. Polarity agreement measures whether VLM and humans assign the same trait direction (above vs. below neutral); chance =50% .
Dimension
Agreement
Chance
Openness (O)
97.6%
50%
Conscientiousness (C)
86.8%
50%
Extraversion (E)
77.3%
50%
Agreeableness (A)
81.2%
50%
Neuroticism (N)
89.1%
50%
Overall
86.6%
50%
Appendix
Table 15: Polarity agreement between VLM and human annotators by OCEAN dimension (chance =50% ). Neutral ratings ( =4 ) are excluded.
PAS
PCS
RS
SCS r
SCS
PAS
1.00
.16
.66
.14
.83
PCS
.16
1.00
.13
.14
.20
RS
.66
.13
1.00
.17
.59
SCS r
.14
.14
.17
1.00
.66
SCS
.83
.20
.59
.66
1.00
Appendix
Table 16: Pearson correlations among per-shot metrics ( n=5,056 ). SCS r = SCS reasoning component. SCS’s high correlation with PAS and SCS r is constructional (weighted composite).
SDXL
StoryDiffusion
FLUX.1-dev
SD3.5-Large
Qwen-2512
Cond
PAS
RS
PAS
RS
PAS
RS
PAS
RS
PAS
RS
C1
0.446
0.429
0.480
0.467
0.464
0.469
0.466
0.493
0.518
0.546
C2
0.392
0.382
0.462
0.449
0.483
0.475
0.452
0.469
0.545
0.560
C3
0.408
0.400
0.471
0.444
0.484
0.460
0.460
0.471
0.540
0.575
C4
0.467
0.423
0.502
0.508
0.522
0.552
0.472
0.488
0.633
0.641
Δ (C4–C1)
+4.8%
–1.4%
+4.6%
+8.8%
+12.6%
+17.7%
+1.5%
–1.0%
+22.1%
+17.4%
Appendix
Table 17: Full ablation study results. C1 = base narrative prompt (plot + appearance); C2 = C1 + raw OCEAN trait tags; C3 = C2 + deduced scene-level emotion labels; C4 (Full TBV) = C1 + explicit visual cues from the TBV lexicon. Appending abstract psychological vocabulary (C2, C3) yields marginal or negative effects, confirming the semantic gap. Only concrete visual grounding (C4) produces consistent gains. n=213 per condition for open-source models.
Model
Comparison
Δ PAS
Wilcoxon p
r
TOST p
Equiv.
FLUX.1-dev
C1 vs C2
+ 0.015
.161
0.13
.009
✓
C1 vs C3
+ 0.020
.372
0.08
.045
✓
C1 vs C4
+ 0.059
.005
0.24
—
—
SDXL
C1 vs C2
− 0.054
.054
− 0.17
.561
×
C1 vs C3
− 0.038
.134
− 0.13
.279
×
C1 vs C4
+ 0.021
.160
0.12
—
—
Appendix
Table 18: Paired Wilcoxon signed-rank tests and TOST equivalence tests ( δ=0.05 ) for PAS across ablation conditions. r = rank-biserial correlation. Equiv. = TOST equivalence at δ=0.05 : ✓ confirmed, × not confirmed, — not applicable. TOST is reported only for C1 vs. C2/C3 comparisons, which are the focus of the equivalence claim; C1 vs. C4 is tested as the intervention condition. ∗ Significant at p<.05 . Closed-source models are evaluated under C1 and C4 only.
Category
N
Mean
SD
Median
Neutral
1053
0.575
0.311
0.655
High-Arousal Positive
952
0.507
0.337
0.571
Low-Arousal Negative
998
0.498
0.338
0.541
High-Arousal Negative
1097
0.477
0.374
0.567
Low-Arousal Positive
981
0.446
0.351
0.474
Appendix
Table 19: PAS distribution across affective scene categories. Neutral scenes obtain the highest mean PAS, while all emotionally salient categories show lower mean PAS.
Model
Direct
Indirect
t
p
SDXL
.430
.423
+0.24
.812
FLUX.1-dev
.497
.467
+1.15
.253
SD3.5-Large
.461
.466
− 0.19
.847
StoryDiffusion
.477
.484
− 0.29
.770
Qwen-2512
.551
.580
− 1.25
.212
GPT Image 2
.590
.601
− 0.32
.749
Appendix
Table 20: Direct vs. indirect PAS comparison (Welch’s t -test). No model shows a significant difference ( p<.05 ).
Figure 4: Qualitative case study across five open-source systems on the same representative story shot. C1 uses only the base narrative prompt; C2 adds abstract OCEAN trait labels; C3 adds a deduced scene-level emotion label; C4 injects explicit TBV-derived visual constraints. Across models, C1–C3 often preserve the broad scene layout but tend to produce generic, weakly visible, or emotion-dominated behavior. C4 more consistently introduces personality-relevant visual evidence, such as exploratory gaze, forward-leaning posture, and open gestures, illustrating the qualitative mechanism behind the PAS and RS gains in Table 2 .
Figure 5: Qualitative case study for the two closed-source APIs. Because GPT Image 2 and Nano Banana perform internal prompt rewriting, we evaluate them only under C1 and C4 conditions. Compared with C1, C4 produces more explicit personality-relevant behavior, including more active body orientation, exploratory engagement with the environment, and less purely scene-driven affect. This qualitative shift is consistent with the quantitative PAS and RS gains reported in Table 2 .
Text-to-image (T2I) models often produce stereotypical images when stereotype-relevant attributes are left unspecified. We ask whether explicit requests can overcome these defaults, especially when the requested attribute contradicts the stereotype. We introduce \mbox{IMPLICIT-Bench}, a challenge benchmark of 1,831 prompt triplets built on a bias knowledge graph (KG) of stereotypes recorded in StereoSet and CrowS-Pairs. Each triplet describes one scene with a neutral prompt, a stereotype prompt, and an anti-stereotype prompt. The neutral prompt measures the model's default, and the two explicit prompts measure whether the model can follow a request on either side of the stereotype. On this benchmark, Qwen-Image depicts the stereotype in 52% of neutral images and the anti-stereotype in 5%. On both Qwen-Image and Stable Diffusion~3, explicit requests that contradict this default succeed far less often than requests that agree with it. Where Qwen-Image's default is strongest, explicitly requesting the stereotype succeeds in 97% of cases, whereas requesting the anti-stereotype succeeds in only 54%. In these scenes, 31% of the images for anti-stereotype requests show the stereotypical attribute instead. We further test whether prompt rewriting and embedding steering can reduce the default on neutral prompts. With the prompt's own KG unit, the best configurations improve a combined bias and alignment score by up to 9.5 points, and with a unit retrieved from the bias KG, by at most 2.9. Debiasing thus depends on knowing which stereotype a prompt evokes.
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.
Recent text-to-image models such as DALLE-3 excel at following diverse prompts yet remain blind to individual aesthetic preferences. We study personalized image generation, where models must align outputs with a user's implicit visual preferences based on a few historically preferred images and a short prompt. To this end, we introduce PIPBench, the first profile-inclusive benchmark for evaluating personalized image generation. We further propose a novel data construction pipeline that leverages psychological and demographic profiling dimensions for both real-user data collection and scalable agent-based data generation. Using PIPBench, we conduct a thorough evaluation of representative line of methods. Our experiments reveal key limitations in existing methods, suggesting new challenges and opportunities for personalized text-to-image synthesis. Project page: https://wuyuhang05.github.io/PIPBench/
Yuhang Wu, Shuxiang Zhang, Wee Hian Ching +2
College of AI, Tsinghua University · Sun Yat-sen University · Shanghai Qi Zhi Institute