HealBench: A Fine-Grained Benchmark for Evaluating Psychological Comfort in Text-to-Image Generation

Introduction

Recently, text-to-image (T2I) generation and instruction-guided image editing have made rapid progress in semantic alignment (Fu et al. 2025b), visual quality, editing fidelity (Yu et al. 2025; Fu et al. 2025a), and human preference modeling (Ma et al. 2025). These advances have enabled generative models to produce images that are increasingly faithful, realistic, and visually appealing. However, as visual generation systems are increasingly used in emotionally sensitive and human-centered scenarios, a more subtle question emerges: can a generated image not only look correct, but also feel right to the viewer? This distinction matters when images are expected to support emotional experience, restoration, or psychological relief rather than merely satisfy semantic or aesthetic criteria. An image may match the prompt and appear visually polished, yet still fail to provide perceived psychological comfort [2].

We argue that this limitation is not merely a weakness of current models, but also a consequence of how the target has been formulated and evaluated. Existing generation pipelines are typically optimized for text-image alignment, perceptual quality, editing fidelity, or generic human preference. These objectives are valuable, but they do not directly capture the affective and restorative experience required by psychological comfort. In particular, current systems suffer from three mismatches. First, there is a target mismatch: perceived psychological comfort depends on affective and restorative cues that are not explicitly captured by existing optimization targets ([32, 2, 25]). Second, there is an evaluation-granularity mismatch: comfort-related failure may arise from different sources, including prompt misalignment, incorrect healing semantics, or poor image quality ([11, 25]). As illustrated in Figure 1, two images may receive similarly low comfort ratings for fundamentally different reasons, making a single scalar score insufficient for diagnostic evaluation. Third, there is a benchmark-and-feedback mismatch: as summarized in Table 1, existing benchmarks and evaluators mainly target preference, alignment, fidelity, compositionality, or generic quality, but do not explicitly target perceived psychological comfort or provide actionable, rationale-bearing feedback. Against this background, we ask: how can perceived psychological comfort be turned into an evaluable and optimization-relevant objective for image generation and image editing? Rather than treating comfort as a vague holistic preference, we formulate healing-oriented generation through three necessary dimensions: text-image alignment, heal accuracy, and image quality.

Benchmark / Evaluator	Venue	Scope	Primary Target	Fine-grained	Comfort-specific	Rationale / Feedback	Optimization
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 ✗
Task-Decomposed AutoEval	ACL’25	T2I evaluator	Automatic multi-criterion quality assessment	Yes ✓	No ✗	Partial	Limited
HealBench (ours)	AAAI’27	T2I + Editing	Perceived psychological comfort	Yes ✓	Yes ✓	Yes ✓	Yes ✓

A generated image should be considered healing only when it satisfies all three dimensions; failure in any one dimension is sufficient to disqualify the output. Accordingly, our goal is not merely to rank images from better to worse, but to identify whether, where, and why a generated image fails to provide perceived psychological comfort.

Based on this formulation, we propose HealBench (shown in Figure 2), a fine-grained benchmark for evaluating the perceived psychological comfort of generated images. HealBench is constructed through a human-and-expert-in-the-loop pipeline that curates high-quality data from a large-scale candidate image-text pool, performs imitative data augmentation, and builds a rigorously filtered test set for unified comparison across models and prompting strategies. Unlike preference-style benchmarks that primarily produce aggregate scores, HealBench is designed around failure-based diagnosis: it identifies whether an output violates text-image alignment, heal accuracy, or image quality, thereby revealing different ways in which a generated image may fail to be healing. Evaluation alone, however, is insufficient for improving generation systems. To make diagnosis actionable, we introduce a score-plus-rationale feedback paradigm, where each result is paired with both a comfort score and an explicit explanation of which dimension fails and why. This structured feedback supports prompt rewriting for both T2I generation and instruction-guided image editing. Since large-scale model iteration cannot rely entirely on costly human evaluation, we further train an automatic evaluator, Tri-dimensional Healing Scorer (THS), on top of HealBench to approximate these structured judgments and study its use as an optimization signal for downstream improvement, including RL-based fine-tuning.

Within this work, we treat psychological comfort as a distinct human-centered objective for visual generation, rather than as a by-product of semantic alignment, aesthetics, or generic preference. Our contributions are as follows: (1) We formulate perceived psychological comfort as a failure-based evaluation problem for healing-oriented visual generation, defined by three necessary dimensions: text-image alignment, heal accuracy, and image quality. (2) We build HealBench, a fine-grained benchmark that diagnoses whether a generated image fails on any of the three dimensions, instead of relying solely on aggregate preference-style scoring. (3) We introduce a score-plus-rationale feedback paradigm and THS, turning structured diagnosis into an actionable signal for prompt refinement and downstream model optimization.

Related Work

Image Generation and Editing Models. Recent visual generation models have made substantial progress in fidelity, semantic alignment, controllability, and human preference alignment. Diffusion models such as DDPM, latent diffusion, Imagen, and DiT enable high-resolution synthesis and strong prompt following ([29, 30, 26] (Ho, Jain, and Abbeel 2020)); autoregressive and token-based models such as DALL·E, Make-A-Scene, and Muse improve complex content modeling and composition ([28, 10, 5]); and flow-matching or rectified-flow methods further enhance training and sampling efficiency ([20, 21, 6]). Preference optimization and reinforcement learning methods further align T2I models with human judgments ([3, 7, 34]), while text-guided editing methods provide instruction-based control over image content and style ([13, 23, 4]). However, these methods mainly optimize whether images are faithful, realistic, editable, or generally preferred. Existing affective methods, such as Affective Image Filter and EmoEdit, focus on explicit emotion cues or targeted emotional editing ([35, 39]), leaving subtler outcomes such as psychological comfort, safety, companionship, and healing underexplored.

Text-to-Image Evaluation Benchmarks and Metrics. Existing T2I evaluation primarily targets semantic correctness, visual quality, compositional reasoning, and generic preference. CLIP-based metrics estimate text-image alignment ([27]); reward and preference models such as PickScore, ImageReward, HPS, and HPSv3 capture overall human preference ([16] (Xu et al. 2023; Wu et al. 2023; Ma et al. 2025)); and aesthetic metrics such as AVA and NIMA assess visual appeal ([24, 33]). Broader benchmarks, including DrawBench, GenEval, T2I-CompBench, VQAScore / GenAI-Bench, and HEIM, evaluate compositionality, spatial relations, reasoning, robustness, safety, and overall quality ([30, 12, 15, 17] (Lin et al. 2024; Li et al. 2024)). Yet these metrics mainly ask whether an image is correct, high-quality, or preferred, rather than whether it is psychologically comforting. Although HRS-Bench introduces emotion-grounded evaluation ([1]), and HEIM and GenAI-Bench cover broad evaluation dimensions ([17] (Li et al. 2024)), deeper affective outcomes such as soothingness and emotional comfort remain insufficiently evaluated. Recent studies further show that metric rankings can be unstable across prompts and protocols ([36]), and that faithfulness metrics remain unreliable on challenging subsets ([31]). HealBench addresses this gap by treating perceived psychological comfort as a dedicated, fine-grained, and failure-diagnostic evaluation target.

Method

Overview

HealBench is constructed in four stages: (1) taxonomy design and web-video-based image collection; (2) captioning, hierarchical labeling, and reward-model-based filtering; (3) category-aware evaluation protocol construction and HealScore annotation; (4) tri-dimensional scorer training and data augmentation. Experts participate in taxonomy design, question revision, and iterative quality checks, while VLMs/LLMs are used as scalable assistants.

Data Collection and Filtering

We first build a two-level taxonomy for psychologically comforting imagery. Experts define 18 coarse-grained categories, supplemented with an Others class, to cover common restorative visual functions. The coarse-grained category set is $\mathcal{T} = \{T_i\}_{i=1}^{19}$. For each coarse category $T_i$, we further define a fine-grained subcategory set $\mathcal{S}_i = \{S_{i,j}\}_{j=1}^{n_i}$, which also includes an Others subtype. In total, the taxonomy contains 19 coarse-grained categories and 102 fine-grained subcategories. Each image $I$ is assigned to exactly one category pair:

Based on this taxonomy, we manually design category-aware retrieval templates to crawl web videos and obtain a raw video set ($\mathcal{D}_{\mathrm{video}}$). We then extract frames from these videos, producing a raw image set ($\mathcal{D}_{\mathrm{raw}}$) of about 100K frames. For each image in $\mathcal{D}_{\mathrm{raw}}$, we use Qwen2.5-VL-7B-Instruct to generate a caption, and then use Claude to assign hierarchical labels. Manual inspection removes invalid or ambiguous category pairs, leaving 92 valid pairs for subsequent experiments. The category pairs and the detailed data collection and annotation process are shown in Figure 3. To remove noisy samples caused by semantic drift, poor composition, and low visual quality, we rank images with four complementary reward signals: CLIP Score, Aesthetic Score, PickScore, and HPSv3. These models serve as coarse quality gates rather than final healing labels. Denote their scores by $r_{\mathrm{clip}}$, $r_{\mathrm{aes}}$, $r_{\mathrm{pick}}$, and $r_{\mathrm{hps}}$, respectively. After independent normalization, we compute the overall score as

Within each fine-grained subcategory, we rank images by $r_{\mathrm{overall}}$ and retain the top 33%, which removes low-quality candidates while preserving category diversity, yielding the filtered real-image subset $\mathcal{D}_{\mathrm{filter}}$.

Evaluation Criteria and HealScore

Existing image-generation metrics mainly emphasize semantic alignment and aesthetic preference, but are insufficient for evaluating psychological soothing effects. We therefore propose HealScore, a structured metric defined on three dimensions: text-image alignment, heal accuracy, and image quality. Here, heal accuracy refers to the perceived healing degree of the image under its target category, not to clinical efficacy. For each category pair $c = (T_i, S_{i,j})$, we build a category-aware question set. An LLM first generates one coarse-level question for the general healing function of $T_i$ and two fine-grained questions for the concrete visual or affective cues of $S_{i,j}$ under each dimension. Experts then revise the questions to remove ambiguous, redundant, or category-mismatched formulations. The resulting question set is

where $q_{d,0}^c$ is the coarse-level question and $q_{d,1}^c, q_{d,2}^c$ are fine-grained supplementary questions. For each image $I \in\mathcal{D}_{\mathrm{filter}}$, an LLM scores every question in $\mathcal{Q}_c(I)$ on a 1–10 scale. Let the resulting scores be $r_{\mathrm{sem},k}(I)$, $r_{\mathrm{heal},k}(I)$, and $r_{\mathrm{qual},k}(I)$. We aggregate them within each dimension:

We report the three dimension scores for diagnosis and use a compact summary score only for ranking, training, and large-scale comparison. Since psychological comfort is the primary objective of the benchmark, the summary score gives the healing dimension twice the weight of the alignment and quality dimensions:

The 1:2:1 design therefore emphasizes the target dimension while retaining the two supporting constraints, rather than masking a failed dimension in diagnostic interpretation. To reduce annotation noise, we manually inspect sampled scored results and iteratively refine the question set by expert feedback. The detailed scoring process of HealScore is shown in Figure 4.

To further enrich the dataset, we select, for each category pair, the top-10 images ranked by HealScore and the top-10 images ranked by the normalized reward-model total score. Their captions are rewritten by Claude and used as prompts for SD3.5-Large to synthesize additional images, producing an augmented set $\mathcal{D}_{\mathrm{augment}}$. We merge $\mathcal{D}_{\mathrm{augment}}$ with $\mathcal{D}_{\mathrm{filter}}$ and remove near-duplicates based on image similarity, obtaining the final dataset $\mathcal{D}_{\mathrm{heal}}$.

Tri-dimensional Heal Scorer (THS)

To approximate rule-based HealScore with a learnable model, we train a Tri-dimensional Heal Scorer (THS). Given an image and its category pair $(T_i, S_{i,j})$, THS predicts three dimension-specific scores for text-image alignment, heal accuracy, and image quality, and combines them into the summary score. We adopt BLIP image-text matching as the backbone. Given image $I$ and category text

we encode image and text separately, project them into a shared embedding space, and obtain normalized embeddings $v$ and $t$. Their fused representation is defined as $h = [v;t;v\odot t]$, where $[\cdot;\cdot]$ denotes concatenation and $\odot$ denotes element-wise multiplication. The fused feature is fed into three independent MLP heads to predict the three dimension scores. Let the supervision targets be $y_{\mathrm{sem}}$, $y_{\mathrm{heal}}$, and $y_{\mathrm{qual}}$, with corresponding predictions $\hat{y}_{\mathrm{sem}}$, $\hat{y}_{\mathrm{heal}}$, and $\hat{y}_{\mathrm{qual}}$. We first optimize the dimension-wise regression loss:

To enforce consistency with the summary HealScore while preserving dimension-wise supervision, we further define

and apply an additional total-score loss: $\mathcal{L}_{\mathrm{total}} = \lVert\hat{y}_{\mathrm{total}}-y_{\mathrm{total}}\rVert^2$. The final objective is $\mathcal{L}=\mathcal{L}_{\mathrm{dim}}+\lambda\mathcal{L}_{\mathrm{total}}$, where $\lambda$ balances the two terms. We optimize THS with AdamW, using different learning rates for the pretrained backbone and the newly initialized heads. During training, we retain only samples with complete annotations on all three dimensions, and select the final model according to the validation error on $\mathcal{L}_{\mathrm{total}}$.

Experiments

Overview

We evaluate HealBench from three perspectives: validity, utility, and scalability, covering benchmark representativeness, tri-dimensional diagnosis, feedback-based improvement, automatic scoring, and downstream optimization.

HealBench-Eval Construction and Validation

Construction To enable controlled evaluation without sacrificing diversity, we construct a compact HealBench-Eval from the filtered real-image subset. For each of the 92 valid

category pairs, we extract SigLIP embeddings, run K-means clustering, and select representative samples. Figure 5 shows that the resulting subset follows the overall HealBench distribution rather than concentrating in a few dense regions, making it suitable for controlled healing-oriented evaluation.

Validation To test whether the tri-dimensional formulation captures a meaningful healing-oriented signal, we compare HealScore-induced rankings with human and VLM preferences on 1,500 image pairs. Annotators, Claude-Haiku, and Claude-Sonnet 4-6 select the more psychologically comforting image. Table 2 shows that human judgments achieve the highest agreement, while Sonnet slightly outperforms Haiku, indicating that HealScore provides a meaningful approximation of this structured preference.

Model	Accuracy	Kendall’s $\tau$	Spearman’s $\rho$
Haiku	0.6831	0.3546	0.4319
Sonnet 4-6	0.6933	0.3868	0.4711
People	0.8012	0.4487	0.5448

Benchmarking Baseline T2I Models

We next benchmark representative diffusion and flow-matching T2I models under the original prompts on healingoriented generation. Table 3 reports standard metrics as reference baselines. Recent models perform strongly on Aesthetic score, CLIPScore, PickScore, and HPSv3, sometimes exceeding the real-image reference on individual metrics. Because these metrics do not target perceived psychological comfort, they should not be treated as evidence of healing-oriented generation quality.

Model	LAION Aesthetic $\uparrow$	CLIPScore $\uparrow$	PickScore $\uparrow$	HPSv3 $\uparrow$
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
SD 3 Medium	4.841	0.965	0.863	11.515
Kolors	4.638	0.901	0.823	9.587
FLUX.1 [dev]	4.798	0.938	0.876	13.585
FLUX.1 [schnell]	4.682	0.943	0.840	10.624
SD 3.5 Large	4.912	0.971	0.867	11.792
SD 3.5 Medium	4.867	0.974	0.867	11.533
sana	4.666	0.967	0.870	12.955
Lumina-Image 2.0	4.950	0.984	0.864	11.919
GT	4.954	0.925	0.838	9.797

Overall healing-oriented performance. Table 4 shows clear gaps across model families. Among original baselines, Playground v2.5 (25.088), PixArt-XL (25.080), and Hunyuan-DiT (25.020) perform best overall, while SDXL Turbo (18.765) and SD 2.1 (19.390) lag behind. Even when advanced generators approach the real-image reference, their scores remain uneven across dimensions and healing functions.

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
Hunyuan-DiT	25.020	6.378	6.178	6.286	24.341	24.849	24.286	28.057	25.427	25.347
Kolors	23.535	6.005	5.866	5.797	23.586	23.924	22.832	26.369	23.267	23.255
Sana-1.6B	24.419	6.386	5.917	6.201	23.748	25.044	24.291	29.624	23.821	24.541
Flow Matching Models
SD 3 Medium	23.465	6.121	5.731	5.882	22.556	23.691	21.617	28.560	23.462	24.780
FLUX.1 [dev]	24.836	6.289	6.152	6.243	24.211	24.195	23.489	29.759	25.424	25.222
FLUX.1 [schnell]	22.717	5.954	5.561	5.642	22.080	21.632	21.108	27.830	23.225	23.913
SD 3.5 Large	24.571	6.346	6.008	6.210	23.694	24.381	23.160	29.057	25.106	25.154
SD 3.5 Medium	24.164	6.321	5.884	6.075	23.285	24.131	22.391	28.730	24.266	25.620
Lumina-Image 2.0	23.169	6.132	5.627	5.782	22.483	22.855	22.035	28.447	23.433	23.602
Prompt-Rewrite Models
Hunyuan-DiT (Rewrite)	26.631 ↑1.612	6.617 ↑0.239	6.711 ↑0.533	6.592 ↑0.306	26.841 ↑2.500	25.843 ↑0.994	26.336 ↑2.050	29.546 ↑1.489	27.183 ↑1.757	25.574 ↑0.226
FLUX.1 [schnell] (Rewrite)	24.514 ↑1.797	6.328 ↑0.374	6.051 ↑0.490	6.084 ↑0.442	24.411 ↑2.331	23.090 ↑1.459	23.942 ↑2.835	28.794 ↑0.965	25.043 ↑1.818	24.552 ↑0.638
SD 3.5 Large (Rewrite)	26.149 ↑1.578	6.638 ↑0.293	6.471 ↑0.464	6.568 ↑0.358	25.851 ↑2.157	24.993 ↑0.611	25.825 ↑2.664	31.057 ↑2.000	26.680 ↑1.575	25.836 ↑0.683
SD 3.5 Medium (Rewrite)	25.598 ↑1.434	6.556 ↑0.235	6.322 ↑0.438	6.398 ↑0.323	25.071 ↑1.786	24.683 ↑0.552	24.554 ↑2.163	29.979 ↑1.248	26.179 ↑1.912	26.046 ↑0.426
Image Editing Models
Pix2Pix (Original)	20.666	5.454	5.081	5.050	20.349	21.401	20.832	23.837	20.832	19.059
Pix2Pix (Rewrite)	22.802 ↑2.136	5.853 ↑0.399	5.664 ↑0.583	5.620 ↑0.570	23.047 ↑2.698	23.193 ↑1.792	23.649 ↑2.817	25.674 ↑1.837	22.719 ↑1.887	20.606 ↑1.547
SDXL-Pix2Pix (Original)	19.307	5.137	4.722	4.727	18.612	19.576	18.351	23.780	19.272	19.733
SDXL-Pix2Pix (Rewrite)	22.053 ↑2.746	5.711 ↑0.574	5.445 ↑0.723	5.453 ↑0.726	22.111 ↑3.499	22.575 ↑2.999	21.464 ↑3.113	26.255 ↑2.475	21.908 ↑2.636	20.917 ↑1.184
Ground Truth
Real Images	24.528	6.375	5.995	6.163	23.186	24.119	25.205	27.901	25.216	24.503

Tri-dimensional analysis. In Table 4, the three dimensions reveal why aggregate scores are insufficient. Strong baselines usually score above 6.0 on Text-Image Alignment and Image Quality, but are lower and more variable on Heal Accuracy: Playground v2.5 obtains 6.479/6.422 on Alignment/Quality but 6.094 on Heal Accuracy, and SD 3.5 Medium obtains 6.321/6.075 but 5.884. Accordingly, perceived healing degree cannot be inferred from alignment or visual quality alone.

Class-level analysis. As shown in Table 4, Class-wise results further show uneven difficulty. Living Companionship is often easiest, reaching 29.624 for Sana-1.6B, 29.759 for FLUX.1 [dev], and 29.057 for SD 3.5 Large. In contrast, Shelter and Safety and Natural Restoration show larger model-dependent variation, and no model dominates all six classes.

HealBench therefore exposes both dimension-specific and function-specific limitations.

Prompt Rewriting as Intervention

We next test whether score-plus-rationale feedback can improve prompts on healing-oriented generation for both text-to-image generation and image editing.

Prompt rewriting improves healing-oriented generation. Table 4 shows that rewriting improves all tested T2I models, including Hunyuan-DiT, FLUX.1 [schnell], SD 3.5 Large, and SD 3.5 Medium, with the most consistent gains on Heal Accuracy.

Prompt rewriting also benefits image editing. For editing, rewritten prompts improve both Pix2Pix and SDXL-Pix2Pix in overall HealScore and all three dimensions.

Qualitative analysis. Figure 6 shows that rewritten prompts produce more coherent, less visually tense images aligned with the intended healing semantics.

Condition	$N$	Improved Ratio (%) $\uparrow$	HAMD-17 $\downarrow$
Random playback	12	45.5	-1.27
Personalized class selection	12	75.0	-3.42
Difference / Gain	–	+29.5	-2.15

User Study on Personalized Healing Delivery

In a preliminary study with 24 college students with subthreshold depression, participants either choose a HEAL class or view randomly assigned healing images. Personalized selection yields a higher improvement ratio (75.0% vs. 45.5%) and larger HAMD-17 reduction (-3.42 vs. -1.27), suggesting that the six-class organization is useful beyond benchmark analysis.

THS as an Automatic Healing Scorer

We next examine whether THS can provide a more suitable automatic evaluation signal for healing-oriented generation than generic scoring models. To this end, we compare THS with several widely used baselines, including PickScore, LAION Aesthetic, CLIPScore, and HPSv3, using the same pairwise ranking protocol as in the preference validation experiment. As shown in Table 6, THS substantially outperforms all generic baselines across all three ranking metrics, achieving 0.8527 in Accuracy, 0.7741 in Kendall’s $\tau$, and 0.9223 in Spearman’s $\rho$. By contrast, the generic scorers remain far behind, with Accuracy values clustered around 0.50–0.56 and much weaker rank correlation. These results indicate that existing scoring models for semantics, aesthetics, or broad human preference do not adequately capture the specific notion of perceived psychological comfort targeted in this work. Results justify the need for a task-specific automatic scorer. Rather than serving only as a descriptive metric, THS provides a scalable evaluation signal that is better aligned with healing-oriented preference and can therefore support both large-scale benchmarking and downstream optimization.

Model	Acc. $\uparrow$	Kendall’s $\tau\uparrow$	Spearman’s $\rho\uparrow$
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

Exploratory RL Fine-Tuning with THS Reward

Finally, we conduct a preliminary study to examine whether THS can serve not only as an evaluator, but also as a reward signal for direct model optimization. To this end, we fine-tune SD 3.5 Medium with Flow-GRPO using a reward defined by a weighted combination of THS and CLIPScore. As shown in Table 7, the Flow-GRPO fine-tuned model improves over the base SD 3.5 Medium across all four metrics, including Overall (23.47 $\rightarrow$ 25.67), Semantic (6.12 $\rightarrow$ 6.48), Heal (5.73 $\rightarrow$ 6.45), and Quality (5.88 $\rightarrow$ 6.28). Notably, the largest gain appears on the Heal dimension, which is consistent with the role of THS as a healing-oriented reward signal. This suggests that the proposed scorer is not only useful for benchmark evaluation, but can also provide an effective optimization signal for improving healing-oriented generation. We emphasize that this experiment is exploratory rather than a fully optimized RL recipe. Nevertheless, the consistent gains across all dimensions provide initial evidence that THS can support downstream alignment toward psychologically comforting image generation.

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

Conclusion

We present HealBench, the first fine-grained benchmark for evaluating perceived psychological comfort in text-to-image generation and image editing. By explicitly defining healing-oriented generation along three complementary dimensions (text-image alignment, heal accuracy, and image quality), we move psychological comfort from an implicit notion of preference to a structured and measurable objective. Building on this formulation, we develop a large-scale human-and-expert-in-the-loop benchmark, introduce the rule-based HealScore, and further propose THS, a learnable automatic evaluator. Our experiments reveal a clear gap between strong performance on conventional generation metrics and true effectiveness under healing-oriented evaluation. Beyond benchmarking, we show that structured score-plus-rationale feedback can directly support generation improvement, and that the HEAL taxonomy provides a useful basis for personalized healing-content delivery. In conclusion, our results establish psychological comfort as a distinct evaluation and optimization target in generative vision, and position HealBench as a strong foundation for future work on affect-aware, human-centered multimodal generation.