Project page · 2026

PortraitAes Intent-Conditioned Structured Portrait Aesthetic Assessment

Junzhou Xie^1,2,‡ Haozhong Xiong^2,‡ Xunyun Tian² Kaile Du¹ Tianchen Yu² Qiang Li² Wei Liu² Jiaming Liu² Ruihua Huang² Yang Shi^2,*,† Guangcan Liu^1,*

¹ Southeast University ² Qwen Business Unit of Alibaba

^‡ These authors contributed equally. · ^* Co-corresponding authors. · ^† Project lead.

Overview of PortraitAes-Bench. — **PortraitAes-Bench.** Each portrait is routed to one of nine photographic intents and evaluated by six specialists spanning 22 secondary criteria.

Abstract

Portrait aesthetic assessment measures how effectively a human-centered image fulfills its photographic intent. Existing methods usually predict one aesthetic score or use general-purpose multimodal models without conditioning on intent. This matters because the same blur, pose, lighting, or framing choice may support one purpose but weaken another.

We introduce PortraitAes-Bench, an 11K-scale benchmark built from expert-authored rubrics for nine photographic intents, six dimensions, and 22 secondary criteria. Following its router–specialist–fusion structure, PortraitAes learns structured judgments with multi-task supervision, Gaussian score calibration, and within-dimension cross-image ranking.

Motivation

Portraits serve functional, commercial, documentary, social, and artistic purposes. A useful assessor must identify that purpose, evaluate visible evidence under explicit criteria, and keep scores comparable across images.

Intent matters

The same visual choice can be appropriate for one photographic intent and harmful for another.

One score is not enough

Structured dimensions reveal where a portrait succeeds or fails instead of hiding the cause in one number.

Ordering matters

Pointwise accuracy alone does not guarantee the correct ranking between visually similar portraits.

Benchmark

PortraitAes-Bench contains 1.6K portraits expanded into approximately 11K linked task records. Each image produces one intent-routing record and six specialist records for structure, composition, image quality, tone, theme and style, and emotion.

Construction pipeline of PortraitAes-Bench. — **Benchmark construction.** Expert rubrics define the router and specialist roles; schema checks, model screening, and expert review precede intent-weighted fusion.

Expert rubrics

Nine intents, six dimensions, 22 criteria, score anchors, and veto or cap rules define what each role evaluates.

Structured records

Records retain scores, semantic labels, subcriteria, visible-evidence rationales, and rule decisions.

Intent-weighted fusion

The final score combines available specialist judgments using the importance weights of the routed intent.

Method

PortraitAes adapts Qwen3.5-VL-9B in two stages. Multi-task supervised fine-tuning first learns the intent router and six rubric-specific specialists. Rank-aligned optimization then improves numerical calibration and cross-image order while retaining structured outputs.

Gaussian calibration and cross-group same-dimension ranking in PortraitAes. — **Rank-aligned optimization.** Gaussian calibration anchors predictions to rubric-derived scores, while same-dimension cross-group ranking penalizes ordinal inversions.

Multi-task supervision

10,286 training images provide approximately 72K examples across the router and six specialist tasks.

Score calibration

A Gaussian reward measures numerical distance for each specialist score and its rubric subcriteria.

Cross-image ranking

Valid examples are compared only within the same dimension and when their reference gap defines a reliable order.

Results

PortraitAes achieves the strongest overall correlations among the evaluated general-purpose multimodal models and specialized aesthetic baselines.

Standard benchmark

0.924PLCC

0.934SRCC

Hard-case set

0.829PLCC

0.795SRCC

Intent-conditioned overall-score distributions for the standard and hard-case evaluations. — **Overall-score distributions.** The standard benchmark spans a broad range, while the hard-case set concentrates visually similar portraits for fine-grained ranking.

Examples

The model predicts an intent-conditioned overall score together with six specialist scores and dimension-level explanations grounded in visible evidence.

Aesthetic portrait · 3.97 / 5

1 / 2