[Paper Review] Fine-Tuning a Large Vision-Language Model for Artwork's Scoring and Critique

Zhehan Zhang, Meihua Qian+5 more|arXiv (Cornell University)|Feb 9, 2026

Aesthetic Perception and Analysis0 citations

TL;DR

The paper fine-tunes a large vision-language model (Qwen2-VL-7B) with multi-task learning to score paintings and generate rubric-aligned critiques, achieving high predictive accuracy and semantically close feedback.

ABSTRACT

Assessing artistic creativity is foundational to creativity research and arts education, yet manual scoring (e.g., Torrance Tests of Creative Thinking) is labor-intensive at scale. Prior machine-learning approaches show promise for visual creativity scoring, but many rely mainly on image features and provide limited or no explanatory feedback. We propose a framework for automated creativity assessment of human paintings by fine-tuning the vision-language model Qwen2-VL-7B with multi-task learning. Our dataset contains 1000 human-created paintings scored on a 1-100 scale and paired with a short human-written description (content or artist explanation). Two expert raters evaluated each work using a five-dimension rubric (originality, color, texture, composition, content) and provided written critiques; we use an 80/20 train-test split. We add a lightweight regression head on the visual encoder output so the model can predict a numerical score and generate rubric-aligned feedback in a single forward pass. By embedding the structured rubric and the artwork description in the system prompt, we constrain the generated text to match the quantitative prediction. Experiments show strong accuracy, achieving Pearson r > 0.97 and MAE about 3.95 on the 100-point scale. Qualitative evaluation indicates the generated feedback is semantically close to expert critiques (average SBERT cosine similarity = 0.798). The proposed approach bridges computer vision and art assessment and offers a scalable tool for creativity research and classroom feedback.

Motivation & Objective

• Motivate scalable automated creativity assessment for paintings to support creativity research and education.
• Develop a multi-task learning framework that jointly predicts scores and generates critiques.
• Leverage structured rubric and artwork descriptions to constrain generated feedback to rubric alignment.

Proposed method

• Fine-tune Qwen2-VL-7B with a lightweight regression head on the visual encoder to predict a 1–100 score.
• Use multi-task learning to enable score prediction and critique generation in a single forward pass.
• Embed the five-dimension rubric (originality, color, texture, composition, content) and artwork description into the system prompt to constrain output.
• Evaluate using an 80/20 train-test split with 1000 paintings and dual expert ratings.
• Measure regression accuracy (Pearson r) and mean absolute error (MAE); assess textual feedback semantics with SBERT cosine similarity (avg 0.798).

Experimental results

Research questions

• RQ1 Can a large vision-language model be fine-tuned to predict quantitative creativity scores for paintings?
• RQ2 Is it possible to generate rubric-aligned critiques in addition to scores in a single forward pass?
• RQ3 Does embedding a structured rubric and description in the prompt constrain generated feedback to rubric alignment?
• RQ4 How well do the model's numeric predictions correlate with expert scores and how close is the generated feedback to expert critiques?

Key findings

• Pearson correlation > 0.97 between model predictions and expert scores on a 100-point scale.
• Mean absolute error (MAE) around 3.95 on the 100-point scale.
• Qualitative feedback is semantically close to expert critiques with SBERT cosine similarity averaging 0.798.
• The approach provides a scalable tool bridging computer vision and art assessment for creativity research and classroom feedback.