Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Defines aesthetic guidance (AG), builds the AesGuide dataset/benchmark of 10,748 photos with scores, analyses and shooting guidance, and proposes Venus, a two-stage SFT framework (AG capability building, then CoT-rationale aesthetic cropping) that achieves SOTA on AG and FLMS cropping.

Problem

General and aesthetic MLLMs give overly positive feedback and lack actionable aesthetic guidance, which also hurts aesthetic cropping; existing cropping models are small, opaque, and non-interactive.

Approach

Introduce an expert-annotated AG dataset with progressive aesthetic questions (overall impression, analysis, guidance) to first instill AG in MLLMs, then fine-tune with CoT aesthetic rationales for both good and bad crops so the MLLM predicts crop boxes with explanations.

Key results

• Venus-L-13B reaches the top AG Mean score of 1.49 on AesGuide, versus 0.98 for GPT-4o and 0.47 for AesExpert.
• Venus-Q achieves IoU 87.01% on FLMS, +1.50 over the prior best specialized model (SAC-Net 85.51) and +15.4 over GPT-4o (71.61).
• Venus-Q reaches recall 92.0% on FLMS versus 67.2% for the best baseline (Qwen-VL-Chat).

Surprising findings

• Aesthetic-oriented MLLM AesExpert scores far below general MLLMs on AG (Mean 0.47), and UNIAA scores only 0.2% recall on FLMS cropping.
• Proprietary GPT-4o and Gemini-2.0-Pro achieve only modest AG scores (0.98 and 1.19 Mean) before Venus fine-tuning.

Limitations

• AG evaluation relies on GPT-based judging validated only on 100 photos by 10 experts.
• Stage-2 cropping evaluated mainly on FLMS; generalization limited to a single out-of-domain benchmark.
• Training data annotation depends on GPT-4o and Qwen2.5-VL-72B generated content.

Experimental setup

Datasets: AesGuide CPC GAIC FCDB FLMS UNIAA auxiliary datasets

Benchmarks: AesGuide FLMS

Metrics: Completeness Preciseness Relevance Mean IoU Disp R (recall IoU>=75%)

Baselines: GPT-4o Gemini-2.0-Pro Qwen-VL-Max AesExpert UNIAA ASM-Net CACNet HCIC SAC-Net UNIC ProCrop Qwen2.5-VL-7B

Builds on: LLaVA CoT UNIAA Q-Bench

Closest papers by meaning

• The Photographer's Eye: Teaching Multimodal Large Language Models to See, and Critique Like Photographers (CVPR 2025 · similarity 0.89)
• Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization (ICLR 2026 · similarity 0.87)
• Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.86)
• Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.86)
• MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 2025 · similarity 0.86)
• Global Compression Commander: Plug-and-Play Inference Acceleration for High-Resolution Large Vision-Language Models (AAAI 2026 · similarity 0.85)