Research map / Image generation / Text-to-image diffusion
ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning
Yuxiang Guo, Jiang Liu, Ze Wang, Hao Chen, Ximeng Sun, Yang Zhao, Jialian Wu, Xiaodong Yu, Zicheng Liu, Emad Barsoum
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 task tokens (<ART>, <MIS>) to the heatmap decoder (ablation) | ImageDoctor heatmap decoder · RichHF-18K | heatmap CC (artifact / misalignment): 0.508 -> 0.532 (artifact CC); 0.123 -> 0.165 (misalignment CC) | Improved |
| add the look step (flaw bounding boxes) (ablation) | ImageDoctor cold start stage 2 · RichHF-18K | misalignment CC: 0.160 -> 0.224 | Improved |
| add the think step (reasoning) (ablation) | ImageDoctor cold start stage 2 · RichHF-18K | average PLCC: 0.708 -> 0.720 | Improved |
| add the grounding reward in reinforcement finetuning (ablation) | ImageDoctor RFT · RichHF-18K | average PLCC: 0.734 -> 0.741 | Improved |
| add reinforcement finetuning after cold start stage 2 (ablation) | ImageDoctor · RichHF-18K | average PLCC: 0.720 -> 0.741 | Improved |
| add heatmap prediction to cold start stage 1 score training (ablation) | ImageDoctor cold start stage 1 · RichHF-18K | average PLCC: 0.660 -> 0.655 | No clear effect |
| run ImageDoctor without the reasoning chain at inference (ablation) | ImageDoctor · RichHF-18K | average PLCC: 0.741 -> 0.729 | Hurt |
| use ImageDoctor scores instead of PickScore as the Flow-GRPO reward | Flow-GRPO on SD3.5-medium · DrawBench | ImageReward: 1.002 -> 1.029 | Improved |
| use ImageDoctor scores instead of RichHF as the Flow-GRPO reward | Flow-GRPO on SD3.5-medium · DrawBench | ImageReward: 0.879 -> 1.029 | Improved |
| add heatmap-based dense pixel-level reward (DenseFlow-GRPO) to scalar ImageDoctor reward | Flow-GRPO with ImageDoctor · DrawBench | ImageReward: 1.029 -> 1.100 | Improved |
| use PickScore as Flow-GRPO reward | Flow-GRPO on SD3.5-medium · DrawBench | CLIPScore: 0.951 -> 0.941 | Hurt |
Closest papers by meaning
- Focus-N-Fix: Region-Aware Fine-Tuning for Text-to-Image Generation (CVPR 2025 · similarity 0.90)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.90)
- Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist Rewards (NeurIPS 2026 · similarity 0.89)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.89)
- Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · similarity 0.88)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.88)
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.
