Research map / Image generation / Text-to-image diffusion
VisualScore: Learning Holistic Visual Quality Scores via Multi-Task Reasoning
Yiting Lu, Fengbin Guan, Yixin Gao, Yan Zhong, Xinge Peng, Jiakang Yuan, Yihao Liu, Bo Zhang, Xin Li, Zhibo Chen, Weisi Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cold-start rejective sampling tuning before GRPO (ablation) | GRPO · EvalMuse | PLCC: 0.640 -> 0.689 | Improved |
| replace binary reward with Gaussian reward (ablation) | VisualScore · EvalMuse | PLCC: 0.689 -> 0.764 | Improved |
| replace binary reward with Gaussian reward (ablation) | VisualScore · AVA | PLCC: 0.737 -> 0.763 | Improved |
| add entropy gating in stage 1 training (ablation) | VisualScore · KADID | PLCC: 0.672 -> 0.645 | Hurt |
| add entropy gating in stage 1 training (ablation) | VisualScore · EvalMuse-50k | PLCC: 0.764 -> 0.742 | Hurt |
| add entropy gating in stage 2 training (ablation) | VisualScore · KADID | PLCC: 0.671 -> 0.735 | Improved |
| add STD filtering on top of entropy gating in stage 2 (ablation) | VisualScore · KADID | PLCC: 0.735 -> 0.745 | Improved |
| add STD filtering on top of entropy gating in stage 2 (ablation) | VisualScore · TAD66k | PLCC: 0.450 -> 0.461 | Improved |
| add entropy gating and STD filtering to multi-task training (ablation) | VisualScore · EvalMuse | PLCC: 0.640 -> 0.764 | Improved |
| train jointly on three tasks instead of single task (with method) (ablation) | VisualScore · EvalMuse | PLCC: 0.731 -> 0.764 | Improved |
| set Gaussian reward sigma to 1 instead of 0.8 (ablation) | GRPO with Gaussian reward · EvalMuse-50k | PLCC: 0.764 -> 0.724 | Hurt |
| use sigma 0.6 instead of 0.8 (ablation) | GRPO with Gaussian reward · EvalMuse-50k | PLCC: 0.764 -> 0.734 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LMM4LMM — LMM4LMM: Benchmarking and Evaluating Large-multimodal Image Generation with LMMs (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.91)
- Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.90)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.90)
- Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing (NeurIPS 2026 · similarity 0.87)
- Think, then Score: Decoupled Reasoning and Scoring for Video Reward Modeling (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
