Research map / Image generation / Text-to-image diffusion
PIPBench: A Profile-Inclusive Framework for Personalized Image Generation Evaluation
Yuhang Wu, Shuxiang Zhang, WEE CHING, Chi Zhang, Miao Liu
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 GPT-5 VLM condition fusion over reference images | Qwen-Image · PIPBench Real-User | Elo: 1427 -> 1765 | Improved |
| add GPT-5 VLM condition fusion over reference images | Qwen-Image · PIPBench Real-User | DIS-R: 12.099 -> 22.160 | Improved |
| use two reference images instead of one in Qwen-Image-Edit | Qwen-Image-Edit · PIPBench Real-User | Elo: 1521 -> 1354 | Hurt |
| add preference conditioning via GPT-5 fusion | Qwen-Image · PIPBench Real-User | CLS-T: 32.677 -> 30.550 | Hurt |
| scale the fusion VLM from QwenVL2.5-7B to 70B | VLM condition fusion · PIPBench Real-User | Elo: 1445 -> 1531 | Improved |
| replace SDXL with Qwen-Image as generation backbone under GPT-5 fusion | GPT-5 fusion · PIPBench Real-User | Elo: 1394 -> 1603 | Improved |
| instruction-tune Qwen2.5-VL-7B on synthetic agent data | VLM caption enrichment · PIPBench Real-User | Elo: 1492 -> 1597 | Improved |
| unfreeze the VLM in the preference compressor (ablation) | preference compression · PIPBench Real-User | Elo: 1318 -> 1534 | Improved |
| use profile-inclusive synthetic data instead of profile-free for VLM tuning (ablation) | Qwen2.5-VL-7B · Real-User | win rate: 47.6% -> 52.4% | Improved |
| use profile-inclusive synthetic data instead of profile-free for VLM tuning (ablation) | Qwen2.5-VL-7B · Real-User | LPIPS-R: 0.7192 -> 0.6901 | Improved |
| use profile-inclusive synthetic data instead of profile-free for VLM tuning (ablation) | Qwen2.5-VL-7B · Real-User | CLS-T: 32.251 -> 31.588 | Hurt |
| add user profile conditioning to synthetic benchmark generation (ablation) | benchmark construction · PIPBench synthetic data | Shannon entropy: 0.9919 -> 0.5811 | Improved |
Closest papers by meaning
- Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting (CVPR 2024 · similarity 0.89)
- DreamBench++: A Human-Aligned Benchmark for Personalized Image Generation (ICLR 2025 · similarity 0.88)
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.87)
- MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 2025 · similarity 0.87)
- MM-R1: Unleashing the Power of Unified Multimodal Large Language Models for Personalized Image Generation (AAAI 2026 · similarity 0.86)
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.
