Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping

Conference: CVPR 2026 arXiv: 2602.23980 Code: https://github.com/PKU-ICST-MIPL/Venus_CVPR2026 Area: Multimodal VLM Keywords: Aesthetic Guidance, Image Cropping, MLLM, Aesthetic Assessment, CoT Reasoning

TL;DR

This paper defines the novel task of Aesthetic Guidance (AG) and constructs the AesGuide benchmark (10,748 photos annotated with aesthetic scores, analyses, and guidance), then proposes Venus, a two-stage framework that first empowers MLLMs with aesthetic guidance capability via progressive aesthetic QA, and subsequently activates aesthetic cropping capability through CoT reasoning, achieving state-of-the-art performance on both tasks.

Background & Motivation

Background: Computational aesthetics has evolved from perception-level tasks such as aesthetic scoring and quality description to higher-level understanding. However, "aesthetic guidance"—identifying aesthetic issues and providing actionable shooting suggestions—remains a critical yet systematically understudied capability.

Limitations of Prior Work: (a) General MLLMs (e.g., GPT-4o) and aesthetic-specialized MLLMs (e.g., AesExpert) tend to produce overly positive evaluations when presented with photos, failing to identify problems or provide actionable recommendations; (b) aesthetic cropping models, while capable of cropping, lack interpretability and interactivity, as they can neither explain their cropping rationale nor adapt to user preferences.

Key Challenge: Existing aesthetic datasets primarily annotate "what is good," lacking guidance-oriented annotations that capture "what is wrong" and "how to improve." Furthermore, MLLMs are misaligned with the human aesthetic reasoning process.

Goal: (a) Construct the first aesthetic guidance dataset and benchmark; (b) empower MLLMs with aesthetic guidance capability; (c) leverage aesthetic guidance capability to activate cropping capability.

Key Insight: Aesthetic guidance follows a human cognitive process of "overall impression → analysis of strengths and weaknesses → improvement suggestions." MLLMs are trained to simulate this process through progressive complexity QA.

Core Idea: Through a two-stage approach—aesthetic guidance empowerment (progressive QA) and aesthetic cropping activation (CoT reasoning rationale)—Venus achieves both aesthetic understanding and aesthetic creation in MLLMs.

Method

Overall Architecture

The framework proceeds in two stages: Stage 1 trains an MLLM on AesGuide to answer aesthetic questions of increasing complexity (scoring → analysis → guidance), establishing aesthetic guidance capability. Stage 2 trains on cropping data using CoT rationales that incorporate aesthetic reasoning (AR), activating cropping capability.

Key Designs

1. AesGuide Dataset Construction (AGGF Framework):

   • Function: Collects 10,748 photos from online platforms and professional photographers, annotated with aesthetic scores, analyses, and guidance.
   • Mechanism: Two-stage annotation pipeline—MLLM refinement (GPT-4o analyzes raw comments → Qwen2.5-VL-72B validates completeness) → expert annotation (20 photography experts review and revise, separating annotations into aesthetic analysis and aesthetic guidance).
   • Design Motivation: Raw online comments are noisy and inconsistent; the MLLM + expert two-stage pipeline ensures annotation quality and consistency.

2. Progressive Aesthetic QA Training:

   • Function: Trains the MLLM to answer aesthetic questions across three levels.
   • Mechanism: Level 1—overall impression ("How is this photo?"); Level 2—detailed analysis ("What compositional issues exist? Is the lighting appropriate?"); Level 3—improvement guidance ("How should it be improved? How to adjust the shooting angle or lighting?").
   • Design Motivation: Simulates the progressive nature of human aesthetic cognition—building intuitive perception before deepening into rational analysis, and finally producing actionable recommendations.

3. Aesthetic Cropping CoT Activation:

   • Function: Trains the model to jointly output crop bounding boxes and aesthetic reasoning.
   • Mechanism: Aesthetic rationales (AR) are generated for both good and poor crops; GPT-4o explains why a cropped region marked with a red bounding box exhibits good or poor composition, and Qwen2.5-VL-72B validates consistency.
   • Design Motivation: Learning crop coordinates alone lacks compositional understanding; CoT rationales compel the model to understand "why crop here," improving both interpretability and interactivity.

4. AG Evaluation Benchmark Design:

   • Function: Uses GPT as a judge to evaluate outputs across three dimensions: completeness, preciseness, and relevance.
   • Mechanism: Each dimension is scored 0–2 with golden annotations as references and GPT-4 as the evaluator. Manual evaluation by 10 experts on 100 samples is conducted to validate the reliability of GPT-based scoring.

Loss & Training

Both stages employ standard instruction fine-tuning: \(\mathcal{L} = -\mathbb{E}\sum_t \log\pi_\theta(y_t|x,q,y_{<t})\). In Stage 1, the visual encoder and connector are frozen, with only the LLM being trained. In Stage 2, the aesthetic guidance MLLM undergoes full-parameter fine-tuning.

Key Experimental Results

Aesthetic Guidance Evaluation (AesGuide Benchmark)

Model	Completeness	Preciseness	Relevance	Mean	Expert
GPT-4o	0.84	1.09	1.01	0.98	1.15
AesExpert-7B	0.33	0.56	0.51	0.47	0.56
UNIAA-7B	1.03	1.02	1.23	1.09	1.01
InternVL 2.5-7B	0.83	1.01	1.02	0.95	0.99
Venus-I (ours)	1.27	1.33	1.81	1.47	1.50
LLaVA-1.5-13B	0.67	0.86	0.41	0.65	0.61
Venus-L-13B (ours)	1.28	1.35	1.83	1.49	1.53

Aesthetic Cropping (FLMS Benchmark)

Model	IoU%↑	Disp↓	Interpretable	Interactive
CACNet	72.8	0.062	✗	✗
TransView	71.5	0.068	✗	✗
GPT-4o	58.3	0.105	✓	✓
Venus-Q (ours)	74.2	0.055	✓	✓

Key Findings

• Venus outperforms GPT-4o by approximately 50% on the aesthetic guidance Mean score (1.47 vs. 0.98), with the largest gain observed in the Relevance dimension (+0.79).
• Aesthetic guidance capability directly benefits cropping performance—omitting Stage 1 and training on cropping directly leads to a significant performance drop.
• A user study involving 1,069 participants shows that 91% desire aesthetic guidance functionality, validating the practical relevance of the task definition.
• Venus achieves simultaneous state-of-the-art cropping performance, interpretability, and interactivity, being the only method to satisfy all three criteria.
• Training with rationales that include "poor crops" yields better results than using only "good crops."

Highlights & Insights

• Contribution of Task Definition: The formal definition of the Aesthetic Guidance (AG) task fills a critical gap in computational aesthetics, with a 91% user survey endorsing its practical demand. This definition is expected to stimulate subsequent research.
• Two-Stage Capability Transfer: The transfer pathway from AG capability to cropping capability is elegant—first enabling the model to "appreciate aesthetics," then enabling it to "create aesthetics," with Stage 1 serving as the foundation for Stage 2. This capability-progressive training paradigm is transferable to other dual "understanding + creation" tasks.
• AGGF Annotation Framework: The annotation pipeline combining MLLM refinement with expert review balances efficiency and quality, offering a practical solution for large-scale subjective task annotation.

Limitations & Future Work

• The AesGuide data primarily originates from online photography communities, which may introduce stylistic bias toward specific aesthetic preferences.
• Cropping is limited to two-dimensional recomposition and does not address richer aesthetic corrections such as 3D perspective adjustment or lighting modification.
• Evaluation relies on GPT as a judge, which may introduce bias for highly subjective aesthetic assessments.
• Personalization is not explored—different users apply different criteria to "good photographs."

Related Work & Insights

• vs. AesExpert: AesExpert focuses on aesthetic perception and description (with a positive bias), whereas Venus targets aesthetic guidance (identifying problems and providing suggestions), representing a fundamentally different positioning.
• vs. CACNet: CACNet is a dedicated cropping model with high IoU but no interpretability; Venus achieves both cropping and explanation simultaneously through CoT rationales.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The AG task definition fills a gap, and the AesGuide dataset is the first of its kind.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Five MLLMs × two tasks, with dual evaluation via GPT and human experts.
• Writing Quality: ⭐⭐⭐⭐ Architecture diagrams are clear; the user study strengthens persuasiveness.
• Value: ⭐⭐⭐⭐⭐ The dataset and benchmark offer high community value, directly targeting practical photography guidance scenarios.

Related Papers

• [CVPR 2026] A3: Towards Advertising Aesthetic Assessment
• [CVPR 2026] See, Hear, and Understand: Benchmarking Audiovisual Human Speech Understanding in Multimodal Large Language Models
• [CVPR 2026] GraphVLM: Benchmarking Vision Language Models for Multimodal Graph Learning
• [CVPR 2026] LFPC: Learning to Focus and Precise Cropping for MLLMs
• [CVPR 2026] Residual Decoding: Mitigating Hallucinations in Large Vision-Language Models via History-Aware Residual Guidance