Research map / Image generation / Text-to-image diffusion
F-Bench: Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration
Lu Liu, Huiyu Duan, Qiang Hu, Liu Yang, Chunlei Cai, Tianxiao Ye, Huayu Liu, Xiaoyun Zhang, Guangtao Zhai
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 the face encoder (ArcFace) (ablation) | F-Eval · FaceQ-Res real-world ID Fidelity | SRCC: 0.603 -> 0.796 | Improved |
| add the projector (ablation) | F-Eval · FaceQ-Res real-world ID Fidelity | SRCC: 0.478 -> 0.796 | Improved |
| use four dimension-wise LoRA experts instead of single LoRA (ablation) | F-Eval · FaceQ-Res real-world ID Fidelity | SRCC: 0.671 -> 0.796 | Improved |
| use four dimension-wise LoRA experts instead of three task-wise LoRA (ablation) | F-Eval · FaceQ-Res real-world ID Fidelity | SRCC: 0.638 -> 0.796 | Improved |
| use four dimension-wise LoRA experts instead of single LoRA (ablation) | F-Eval · FaceQ-Gen Quality | SRCC: 0.793 -> 0.849 | Improved |
| add the face encoder (ablation) | F-Eval · FaceQ-Gen Quality | SRCC: 0.815 -> 0.849 | Improved |
| add the projector (ablation) | F-Eval · FaceQ-Gen Quality | SRCC: 0.808 -> 0.849 | Improved |
| add face encoder on face generation correspondence (ablation) | F-Eval · FaceQ-Gen Correspondence | SRCC: 0.835 -> 0.847 | No clear effect |
| use dimension LoRA instead of single LoRA on customization quality (ablation) | F-Eval · FaceQ-Cus Quality | SRCC: 0.941 -> 0.946 | No clear effect |
| fine-tune MINT-IQA on FaceQ | MINT-IQA · FaceQ-Gen Quality | SRCC: 0.7002 -> 0.8312 | Improved |
| replace best baseline with F-Eval on customization ID fidelity | PromptIQA (fine-tuned) · FaceQ-Cus ID Fidelity | SRCC: 0.8693 -> 0.9188 | Improved |
| replace best baseline with F-Eval on real-world restoration quality | MANIQA (fine-tuned) · FaceQ-Res real-world Quality | SRCC: 0.8356 -> 0.8448 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ImageReward — ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023) · better on 2 of 2 shared results
- DSL-FIQA — DSL-FIQA: Assessing Facial Image Quality via Dual-Set Degradation Learning and Landmark-Guided Transformer (CVPR 2024) · 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
- LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models (NeurIPS 2026 · similarity 0.88)
- AI-Face: A Million-Scale Demographically Annotated AI-Generated Face Dataset and Fairness Benchmark (CVPR 2025 · similarity 0.88)
- FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs (CVPR 2025 · similarity 0.87)
- MultiHuman-Testbench: Benchmarking Image Generation for Multiple Humans (NeurIPS 2025 · similarity 0.87)
- Limitations of Face Image Generation (AAAI 2024 · similarity 0.87)
- A-Bench: Are LMMs Masters at Evaluating AI-generated Images? (ICLR 2025 · 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.
