PortraitAes: Intent-Conditioned Structured Portrait Aesthetics Assessment

Junzhou Xie; Haozhong Xiong; Xunyun Tian; Kaile Du; Tianchen Yu; Qiang Li; Wei Liu; Jiaming Liu

cs.CVOct 4, 2026

PortraitAes: Intent-Conditioned Structured Portrait Aesthetics Assessment

Authors: Junzhou Xie, Haozhong Xiong, Xunyun Tian, Kaile Du, Tianchen Yu, Qiang Li, Wei Liu, Jiaming Liu, +3 more

Organizations: Southeast University · Qwen Business Unit of Alibaba

Abstract

Portrait aesthetic assessment assigns comparable scores according to how effectively human-centered images fulfill their photographic intent. These scores support data filtering, candidate selection, and preference modeling in image-generation pipelines. Existing methods typically predict a single aesthetic score or use general-purpose MLLMs without conditioning on photographic intent. This omission matters because the same blur, pose, lighting, or framing choice may serve one photographic intent but undermine another. These models thus learn context-agnostic aesthetic priors and yield inconsistent, inaccurate, misleading judgments for portraits with distinct photographic objectives. We introduce PortraitAes-Bench, an 11K-scale benchmark that decomposes this task into intent-conditioned subjudgments. Expert-authored rubrics define nine photographic intents, six first-level dimensions, and 22 secondary criteria. They support a structured pipeline for intent routing, specialist assessment, verification, and score fusion. Following this structure, we train PortraitAes with multi-task supervision. We then improve score comparability through Gaussian score calibration and within-dimension cross-image ranking. On the standard benchmark, PortraitAes achieves a Pearson correlation of 0.924 and a Spearman rank correlation of 0.934. On the hard-case set, its Pearson correlation is 0.829 and its Spearman rank correlation is 0.795. Across both sets, PortraitAes outperforms the evaluated general-purpose MLLMs and specialized aesthetic baselines.

Figures & tables

Figure 1: Overview of PortraitAes-Bench. Each portrait is routed to one of nine photographic intents and evaluated by six specialists spanning 22 secondary criteria, yielding structured routing, rating, and intent-conditioned fusion targets.

Figure 2: Construction of PortraitAes-Bench. Expert rubrics define an intent router and six specialist roles; schema checks, Qwen3.5 Flash screening, and expert review precede intent-weighted fusion.

Figure 3: Intent-conditioned overall-score distributions. The standard benchmark spans a broad range, whereas the hard-case set concentrates visually similar portraits for fine-grained ranking.

Figure 4: Illustration of the numerical rewards. Gaussian calibration anchors predictions to rubric-derived scores; same-dimension cross-group ranking penalizes ordinal inversions. Images and rollout values illustrate the mechanism rather than report an evaluation case.

Model	Structure		Composition		Quality		Tone		Theme		Emotion		Overall
Model	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
Qwen3.5-VL-9B	0.833	0.822	0.724	0.742	0.741	0.698	0.807	0.789	0.706	0.714	0.698	0.740	0.857	0.874
InternVL3-8B	0.393	0.453	0.518	0.505	0.581	0.623	0.530	0.570	0.448	0.511	0.297	0.296	0.607	0.668
PortraitCraft-4B †	0.698	0.801	0.647	0.760	0.738	0.773	0.578	0.698	0.613	0.754	0.535	0.655	0.652	0.785
Qwen3.5 Flash	0.791	0.837	0.620	0.700	0.807	0.798	0.795	0.818	0.590	0.705	0.667	0.723	0.804	0.763
ArtiMuse †	0.741	0.785	0.744	0.796	0.556	0.602	0.688	0.717	0.725	0.756	0.712	0.703	0.785	0.804

Table 1: Standard PortraitAes-Bench results (PLCC/SRCC; higher is better).

\dagger

denotes diagnostic correlations from a holistic-only scorer.

Model	Structure		Composition		Quality		Tone		Theme		Emotion		Overall
Model	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
Qwen3.5-VL-9B	0.515	0.513	0.346	0.281	0.558	0.506	0.602	0.574	0.385	0.474	0.391	0.413	0.649	0.629
InternVL3-8B	0.030	0.036	0.156	0.112	0.402	0.476	0.228	0.273	-0.010	-0.033	0.121	0.184	0.210	0.215
PortraitCraft-4B †	0.406	0.403	0.361	0.331	0.479	0.463	0.151	0.145	0.342	0.328	0.288	0.297	0.406	0.400
Qwen3.5 Flash	0.522	0.522	0.240	0.214	0.578	0.543	0.624	0.601	0.304	0.411	0.407	0.400	0.615	0.635
ArtiMuse †	0.430	0.421	0.407	0.398	0.336	0.346	0.368	0.350	0.410	0.418	0.387	0.383	0.501	0.495

Table 2: PortraitAes-Bench hard-case results (PLCC/SRCC; higher is better).

\dagger

denotes diagnostic correlations from a holistic-only scorer.

Method	Training component			PLCC/SRCC
Method	SFT	Gaussian	Cross-BPR	Structure	Composition	Quality	Tone	Theme	Emotion	Overall
Qwen3.5-VL-9B				0.833/0.822	0.724/0.742	0.741/0.698	0.807/0.789	0.706/0.714	0.698/0.740	0.857/0.874
PortraitAes-SFT	$\checkmark$			0.856/0.896	0.852/0.885	0.873 / 0.868	0.904/0.905	0.855 /0.877	0.810/0.830	0.909/0.917
PortraitAes-GRPO (Gaussian only)	$\checkmark$	$\checkmark$		0.863 / 0.901	0.869 / 0.897	0.872/ 0.868	0.907 / 0.907	0.854/ 0.883	0.820 / 0.837	0.915 / 0.924
PortraitAes-GRPO	$\checkmark$	$\checkmark$	$\checkmark$	0.875/0.912	0.878/0.903	0.889/0.891	0.914/0.913	0.879/0.901	0.847/0.863	0.924/0.934

Table 3: Component ablation on the standard PortraitAes-Bench. Each result is reported as PLCC/SRCC; higher is better. Checkmarks denote enabled training components.

Method	Training component			PLCC/SRCC
Method	SFT	Gaussian	Cross-BPR	Structure	Composition	Quality	Tone	Theme	Emotion	Overall
Qwen3.5-VL-9B				0.515/0.513	0.346/0.281	0.558/0.506	0.602/0.574	0.385/0.474	0.391/0.413	0.649/0.629
PortraitAes-SFT	$\checkmark$			0.562/ 0.668	0.558/0.524	0.737 / 0.723	0.795 / 0.753	0.570/ 0.598	0.549/0.559	0.780/ 0.758
PortraitAes-GRPO (Gaussian only)	$\checkmark$	$\checkmark$		0.617 /0.624	0.597 / 0.540	0.713/0.694	0.768/0.731	0.661 /0.590	0.610 / 0.587	0.789 /0.755
PortraitAes-GRPO	$\checkmark$	$\checkmark$	$\checkmark$	0.679 / 0.650	0.646 / 0.614	0.768 / 0.750	0.789 / 0.752	0.689 / 0.627	0.650 / 0.629	0.829 / 0.795

Table 4: Component ablation on the PortraitAes-Bench hard-case set. Each result is reported as PLCC/SRCC; higher is better. Checkmarks denote enabled training components.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Resource	Scale	Port.	Supervision	Dim.	Text
AVA ( Murray et al., 2012 )	255K	no	ratings	global	limited
AADB ( Kong et al., 2016 )	10K	no	score/attr./rank	attr.	no
ArtiMuse ( Cao et al., 2026 )	10K	no	score/attr./text	8 attr.	yes
UniPercept ( Cao et al., 2025 )	800K	no	public VR/VQA	perception	yes
Qwen-Image-Bench ( Li et al., 2026 )	1K	no	judge scores	5D/56F	yes
MPAD/MPAAD ( Huang et al., 2022 )	18K/3K	yes	overall/attr.	3 attr.	no

Appendix

Table 5: Comparison with representative aesthetics and portrait-related resources. PortraitAes-Bench emphasizes intent-conditioned, multi-dimensional portrait understanding rather than a single global score.

Figure 5: Specialist-score distributions in PortraitAes-Bench. The six portrait-aesthetic dimensions show distinct score profiles rather than a single shared difficulty pattern, supporting dimension-wise evaluation of structure, composition, image quality, tone, theme/style, and emotion.

Intent	Struct.	Comp.	Quality	Tone	Theme	Emotion
I Functional	0.20	0.20	0.30	0.15	0.10	0.05
II Commercial	0.10	0.15	0.35	0.25	0.10	0.05
III Narrative	0.15	0.15	0.15	0.20	0.10	0.25
IV Aesthetic	0.10	0.15	0.15	0.30	0.20	0.10
V Social	0.05	0.15	0.15	0.20	0.30	0.15
VI Artistic	0.10	0.15	0.05	0.20	0.20	0.30

Appendix

Table 6: Intent-conditioned fusion weights used by the router.

Hyperparameter	Value	Hyperparameter	Value
Completions per prompt $K$	8	Training epochs	1
Learning rate	$2\times 10^{-6}$	Warmup ratio	0.05
KL coefficient	0.02	Clip radius	0.12
Per-device batch size	4	Gradient accumulation	1
Calibration/ranking weights	1.4 / 0.8	Semantic/format weights	0.05 / 0.02
Gaussian bandwidth $\sigma$	0.5	Primary/sub-score weights	0.8 / 0.2

Appendix

Table 7: Hyperparameters for rank-aligned GRPO.

Figure 6: Structured predictions for twelve portraits across six intent categories, grouped by category and ordered by decreasing overall score. Colored bars show the six specialist scores.

Figure 7: Dimension-level aesthetic explanations for an aesthetic portrait with an overall score of 3.97.

Figure 8: Dimension-level aesthetic explanations for a snapshot with an overall score of 2.61.

Explore similar work

Preferences Order, Ratings Anchor: From Fused Expert Aesthetic Ground Truth to Self-Distillation
May 19, 2026Yuanpei Zhao, Jie Lin, Chao Zhang +5Aesthetics Large-Scale Image Datasets
1Sichuan University · 2NetEase Fuxi AI Lab · †Work done during an internship at NetEase Fuxi AI Lab. +1
Visual Aesthetic Benchmark: Can Frontier Models Judge Beauty?
May 12, 2026Yichen Feng, Yuetai Li, Chunjiang Liu +14Aesthetics Evaluation Benchmarks
Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation
May 27, 2026Niantong Li, Guangzheng Hu, Weixu Qiao +35Creativity Fidelity
Alibaba Inc.

May 19, 2026, cs.CVJ/K move · Enter open · S save

Preferences Order, Ratings Anchor: From Fused Expert Aesthetic Ground Truth to Self-Distillation

Yuanpei Zhao, Jie Lin, Chao Zhang, Yilin Wang+4

1Sichuan University · 2NetEase Fuxi AI Lab · †Work done during an internship at NetEase Fuxi AI Lab.+1

Aesthetics Large-Scale Image Datasets Rater Human Annotations Art Ground Truth+2

Open paper AskPDF arXiv

Pairwise preferences and pointwise ratings are the two dominant annotation protocols in image aesthetic assessment (IAA), yet existing benchmarks adopt only one, leaving their complementarity unmeasured under controlled conditions. We introduce PPaint, a matched dual-protocol benchmark in which 15 domain experts, 5 per category, annotate 150 Chinese paintings under both protocols across five aesthetic dimensions, collecting 45,900 pairwise expert judgments through a locally dense preference design alongside the matched ratings. The matched design reveals complementary strengths: preferences yield more consistent ordinal rankings, while ratings anchor the absolute score scale. Fusing both signals via two independent preference-to-score methods yields a fused expert ground truth on which the two constructions converge to nearly identical scores. The same preference-to-score principle extends to label-free VLM training. PSDistill converts VLM pairwise judgments into calibrated pseudo-scores via an Elo reference pool, and trains the same VLM with confidence-weighted ranking optimization to produce a single-pass aesthetic scorer. Trained on a single painting category, the distilled Qwen3-VL-8B improves mean SRCC from 0.504 to 0.709 across all three categories, outperforming all open-source baselines including the dedicated aesthetic model ArtiMuse and matching closed-source Gemini-3.1-Pro within 0.04 SRCC at single-pass inference cost, with cross-domain transfer further validated on APDDv2. We will release the full PPaint dataset and training code.

Abstract

Explore similar work

Preferences Order, Ratings Anchor: From Fused Expert Aesthetic Ground Truth to Self-Distillation

Visual Aesthetic Benchmark: Can Frontier Models Judge Beauty?

Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation