HanMoVLM: Large Vision-Language Models for Professional Artistic Painting Evaluation

Hongji Yang; Yucheng Zhou; Wencheng Han; Songlian Li; Xiaotong Zhao; Jianbing Shen

cs.CVMar 11, 2026

HanMoVLM: Large Vision-Language Models for Professional Artistic Painting Evaluation

Authors: Hongji Yang, Yucheng Zhou, Wencheng Han, Songlian Li, Xiaotong Zhao, Jianbing Shen

Organizations: SKL-IOTSC, CIS, University of Macau · CSE, Shandong University · Online-Video BU, Tencent

Abstract

While Large Vision-Language Models (VLMs) demonstrate impressive general visual capabilities, they remain artistically blind and unable to offer professional evaluation of artworks within specific artistic domains like human experts. To bridge this gap, we transform VLMs into experts capable of professional-grade painting evaluation in the Chinese Artistic Domain, which is more abstract and demands extensive artistic training for evaluation. We introduce HanMo-Bench, a new dataset that features authentic auction-grade masterpieces and AI-generated works, grounded in real-world market valuations. To realize the rigorous judgment, we propose the HanMoVLM and construct a Chain-of-Thought (CoT) validated by experts. This CoT guides the model to perform expert-level reasoning: from content identification and Region of Interest (RoI) localization to professional evaluation, guided by both theme-specific evaluation and typical three-tier evaluation in Chinese paintings. Furthermore, we design a reward function to refine the reasoning process of the HanMoVLM to improve the accuracy. We demonstrate that HanMoVLM can serve as a critical backbone for Test-time Scaling in image generation. By acting as a high-quality verifier, HanMoVLM enables generative models to select the most artistically superior outputs from multiple candidates. Experimental results and human studies confirm that the proposed HanMoVLM effectively bridges the gap, achieving a high consistency with professional experts and significantly improving the quality of Chinese Painting generation.

Figures & tables

Figure 1: The motivation of our study. (a) Generic VLMs suffer from artistic misalignment, evaluation gaps, and low-quality data when handling domain-specific artistic content such as Chinese paintings; (b) The expert-level VLMs introduce the structured expert chain-of-thought, to produce reliable final scores; (c) Experimental results show that our CoT and RL significantly improve performance.

Figure 2: The overall framework of HanMoVLM. For visual artistic understanding, we finetune HanMoVLM with expert-level CoT and expert answer via SFT, and then perform RFT based on the expert reward function. For visual artistic generation, we apply the frozen HanMoVLM as the external verifier to evaluate the images sampled from existing T2I models.

Figure 3: The expert-level Chain-of-Thought in our HanMoVLM. When the generic VLM evaluates Chinese paintings, the model tends to generate outputs based on general knowledge (non-expert) rather than following professional guidelines. As a result, its analysis does not reflect the experts’ procedures, leading to a decline in terms of professional level, reliability, hallucination, and human logic.

Figure 5: The construction pipeline of HanMo-Bench.

Proprietary Models
Model	Size	$\text{MAE}\downarrow$	$\text{RMSE}\downarrow$	$\text{Accuracy(\%)}\uparrow$	$\text{BERTScore}\uparrow$
Gemini-3-Flash [ 13 ]	-	0.9883	1.3472	31.17	–
+ Thinking	-	0.9633	1.4355	34.83	0.3052
+ Expert CoT	-	0.9466	1.3351	37.00	0.3035
Gemini-3-Pro [ 13 ]	-	0.9866	1.3589	33.83	–
+ Thinking	-	0.9783	1.3311	35.17	0.2855

Table 1: The comparison of our HanMoVLM with other models. 10 and 10 indicate the best and second best results, respectively.

Model	Size	$\text{mIoU}\uparrow$	$\text{BERTScore}({\text{RoI}})\uparrow$	$\text{Theme}\uparrow$
Qwen3-VL-Instruct	4B	0.0821	0.2852	0.7346
HanMoVLM (SFT)	4B	0.2708	0.4963	0.8305
HanMoVLM (RFT)	4B	0.2917	0.4911	0.8411
Qwen3-VL-Instruct	8B	0.1401	0.3214	0.7218
HanMoVLM (SFT)	8B	0.2835	0.4872	0.8355
HanMoVLM (RFT)	8B	0.3193	0.5126	0.8515

Table 2: The comparison in mIoU and Theme identification of our VLM with other models.

Figure 6: Generated results in different ratings groups (zoom in for details).

Figure 7: Generated results with the same prompt across different ratings.

$R_{\text{acc}}$	$R_{\text{BERT}}$	$R_{\text{mIoU}}$	Accuracy (%) $\uparrow$	BERTScore $\uparrow$	mIoU $\uparrow$
✓		✓	42.18	0.5529	0.2914
✓	✓		42.36	0.5492	0.2728
	✓	✓	39.82	0.5562	0.2902
✓	✓	✓	42.50	0.5568	0.2917

Table 3: The ablation study of using different rewards.

Figure 12: The case of RoI in HanMoVLM and the corresponding predicted description and gt description in HanMoVLM. Zoom in for more details.

Rating Type	Top-1 Accuracy $\uparrow$	Accuracy $\uparrow$	Kendall’s $\tau$ $\uparrow$	Spearman’s $\rho$ $\uparrow$
Aesthetic Score	0.045	0.125	-0.258	-0.282
PickScore [ 20 ]	0.409	0.341	0.106	0.073
HPSv2.1 [ 42 ]	0.364	0.193	-0.151	-0.173
GPT-4	0.227	0.341	0.197	0.218
GPT-5.1	0.273	0.364	0.121	0.209
HanMoVLM-8B	0.818	0.682	0.758	0.845

Table 4: Comparison of ranking accuracy in different models.

Figure 13: Visual results grouped by the score from our HanMoVLM. Zoom in for details.

Vision-Language Model-as-a-Judge

Vision-language model-as-a-judge uses multimodal models to assess images, video, and visual-agent outputs against instructions, rubrics, or human preferences.

See all 55 papers

CArtBench: Evaluating Vision-Language Models on Chinese Art Understanding, Interpretation, and Authenticity
Apr 13, 2026Xuefeng Wei, Zhixuan Wang, Xuan Zhou +5VLM Evaluation VLM Reasoning
Nara Institute of Science and Technology · Liaoning Normal University
Preferences Order, Ratings Anchor: From Fused Expert Aesthetic Ground Truth to Self-Distillation
May 19, 2026Yuanpei Zhao, Jie Lin, Chao Zhang +5Pairwise Preference Learning Personalized Image Aesthetics Assessment
1Sichuan University · 2NetEase Fuxi AI Lab · †Work done during an internship at NetEase Fuxi AI Lab. +1
SalArt-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images
Jun 10, 2026Ruotian Zhang, Xiaoxiao Sun, Junzhe Huang +2VLM Evaluation VLM Hallucination
Zhejiang University · The University of Queensland

Apr 13, 2026J/K move · Enter open · S save

CArtBench: Evaluating Vision-Language Models on Chinese Art Understanding, Interpretation, and Authenticity

Xuefeng Wei, Zhixuan Wang, Xuan Zhou, Zhi Qu+4

Nara Institute of Science and Technology · Liaoning Normal University

VLM Evaluation VLM Reasoning Cultural Heritage

Open paper AskPDF arXiv

We introduce CARTBENCH, a museum-grounded benchmark for evaluating vision-language models (VLMs) on Chinese artworks beyond short-form recognition and QA. CARTBENCH comprises four subtasks: CURATORQA for evidence-grounded recognition and reasoning, CATALOGCAPTION for structured four-section expert-style appreciation, REINTERPRET for defensible reinterpretation with expert ratings, and CONNOISSEURPAIRS for diagnostic authenticity discrimination under visually similar confounds. CARTBENCH is built by aligning image-bearing Palace Museum objects from Wikidata with authoritative catalog pages, spanning five art categories across multiple dynasties. Across nine representative VLMs, we find that high overall CURATORQA accuracy can mask sharp drops on hard evidence linking and style-to-period inference; long-form appreciation remains far from expert references; and authenticity-oriented diagnostic discrimination stays near chance, underscoring the difficulty of connoisseur-level reasoning for current models.

Abstract

Vision-Language Model-as-a-Judge

Explore similar work

CArtBench: Evaluating Vision-Language Models on Chinese Art Understanding, Interpretation, and Authenticity

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

SalArt-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images