Despite advances in photorealism, semantic fidelity, and aesthetics, text-to-image generation and image editing remain understudied in a human-centered dimension: whether their outputs feel psychologically comforting. Grounded in cognitive evaluation theory, we define psychological comfort as an affective response shaped by viewers' appraisal of restorative and emotionally supportive visual cues. We introduce HealBench, a fine-grained benchmark of 30,000 high-quality image-text pairs spanning 19 categories and 92 subcategories, selected from 100,000 candidates through a human- and expert-in-the-loop pipeline combining vision-language model matching, human preference models, expert filtering, and exemplar-guided augmentation. We also construct HealBench-Eval, a feature-clustered evaluation set of 1,095 instances. Using HealBench-Eval, we conduct objective and subjective evaluations of 14 representative image generation models and prompt-rewriting strategies. Results reveal a gap between conventional visual quality and perceived comfort: strong semantic alignment and aesthetics do not guarantee emotionally restorative images. Reinforcement learning experiments further show that HealScorer can serve as a reward signal for assessing and improving the emotional and restorative quality of generated images. These findings establish psychological comfort as a distinct dimension of human-centered image generation and position HealBench as a foundation for affect-aware generative models.
Figures & tables
Figure 1: Despite sharing a prompt class, the left image scores poorly on general models (like CLIP) but high in psychological comfort (HealScore). Conversely, the right image has high evaluation scores but low comfort.
Benchmark / Evaluator
Venue
Scope
Primary Target
Fine- grained
Comfort- specific
Rationale / Feedback
Optimi- zation
Pick-a-Pic
NeurIPS’23
T2I
Pairwise human preference
No ✗
No ✗
No ✗
Limited
AIGCIQA2023
LNCS’24
T2I
Quality, authenticity, and correspondence
Partial
No ✗
No ✗
No ✗
T2I-CompBench
NeurIPS’23
T2I
Compositional correctness
Yes ✓
No ✗
No ✗
No ✗
HATIE
CVPR’25
Editing
Human-aligned editing quality
Yes ✓
No ✗
No ✗
Limited
NTIRE 2025 T2I QA
CVPR’25
T2I
Fine-grained alignment and structural distortion
Yes ✓
No ✗
Partial
No ✗
ICE-Bench
ICCV’25
T2I + Editing
Unified evaluation for creation and editing
Yes ✓
No ✗
No ✗
No ✗
Table 1: Comparison with representative benchmarks and evaluators for T2I generation and image editing. Existing resources mainly focus on preference, alignment, fidelity, compositionality, or generic quality assessment, whereas HealBench explicitly targets perceived psychological comfort with fine-grained diagnostic dimensions and actionable rationale-bearing feedback .
Figure 2: Overview of HealBench. HealBench organizes healing-related visual content into 19 categories and 92 subcategories, which are further grouped into six major classes according to different healing functions. For each category, we present both real-world examples and enhanced images generated via SD 3.5-L, which is also used for dataset augmentation.
Figure 3: Overview of the HealBench construction pipeline. The benchmark is built through three stages: data collection, reward-model-based filtering and augmentation, and tri-dimensional evaluation-criteria generation. Human and expert review is incorporated throughout the pipeline to ensure the quality of both data instances and evaluation criteria.
Figure 4: Overview of the two-stage HealScore pipeline. HealScore first evaluates each image with three category-level questions on text-image alignment , heal accuracy , and image quality , and then refines the judgment with six subcategory-level questions. The final score is computed by weighted aggregation of the resulting multi-dimensional responses.
Figure 5: t-SNE visualization of SigLIP embeddings for the full HealBench data and the selected HealBench-Eval.
Model
Accuracy
Kendall’s τ
Spearman’s ρ
Haiku
0.6831
0.3546
0.4319
Sonnet 4-6
0.6933
0.3868
0.4711
People
0.8012
0.4487
0.5448
Table 2: Agreement between HealScore-induced rankings and pairwise preferences from human annotators and VLM-based judges.
Model
LAION Aesthetic ↑
CLIPScore ↑
PickScore ↑
HPSv3 ↑
SD 2.1
4.403
0.885
0.785
3.406
SDXL Base 1.0
4.631
0.885
0.790
3.602
pixArt-XL
4.717
0.946
0.864
12.623
SDXL Turbo
4.095
0.975
0.812
5.543
Playground v2.5
5.164
0.988
0.899
12.712
Hunyuan-Dit
5.271
0.976
0.872
10.991
Table 3: Quantitative comparison of different text-to-image models across multiple evaluation metrics.
Model
Overall
Text-Image Alignment
Heal Accuracy
Image Quality
Natural Restoration
Atmospheric Immersion
Shelter and Safety
Living Companionship
Rhythmic Focus
Awe and Transcendence
Diffusion Models
SD 2.1
19.390
5.172
4.820
4.578
18.754
19.678
18.484
21.780
19.412
20.147
SDXL Base 1.0
20.723
5.319
5.173
5.057
20.258
21.269
21.915
24.284
19.859
20.481
PixArt-XL
25.080
6.435
6.169
6.308
24.953
24.832
25.115
27.787
25.036
24.844
SDXL Turbo
18.765
5.419
4.517
4.312
17.372
18.689
17.536
19.837
20.549
18.695
Playground v2.5
25.088
6.479
6.094
6.422
24.551
25.304
24.449
28.177
25.162
25.253
Table 4: Baseline summary table. All metrics are positive indicators. For prompt-rewrite models, the orange mark denotes the improvement over the corresponding non-rewrite base model. For image editing models, the blue mark denotes the improvement over the corresponding original model.
Figure 6: Qualitative examples of our prompt rewriting strategy. Using the same image generation model, targeted prompt rewriting markedly improves generation outcomes, yielding more coherent and safer images with substantially higher HealScores across multiple categories.
Condition
N
Improved Ratio (%) ↑
HAMD-17 ↓
Random playback
12
45.5
-1.27
Personalized class selection
12
75.0
-3.42
Difference / Gain
–
+29.5
-2.15
Table 5: User study on personalized healing delivery with 24 subthreshold-depression college students. Participants in the personalized condition could choose among the six HEAL classes, while those in the random condition viewed randomly assigned healing images. Mood change was measured with HAMD-17 after one hour of viewing.
Model
Acc. ↑
Kendall’s τ↑
Spearman’s ρ↑
PickScore
0.5593
0.1525
0.2254
LAION Aesthetic
0.5507
0.1077
0.1595
CLIPScore
0.5413
0.1272
0.1882
HPSv3
0.5013
0.0379
0.0595
THS
0.8527
0.7741
0.9223
Table 6: THS vs. generic scorers on pairwise healing-preference ranking.
Model
Overall
Semantic
Heal
Quality
SD 3.5 Medi (base)
24.16
6.32
5.88
6.08
SD 3.5 Medi + Flow-GRPO
25.67
6.48
6.45
6.28
Table 7: Main results of the base SD 3.5 Medium model and its Flow-GRPO fine-tuned variant. Higher values are highlighted in blue.