Research map / Image generation / Text-to-image diffusion
Think in Strokes, Not Pixels: Process-Driven Image Generation via Interleaved Reasoning
Lei Zhang, Junjiao Tian, Zhipeng Fan, Kunpeng Li, Jialiang Wang, Weifeng Chen, Markos Georgopoulos, Felix Juefei-Xu, Julian McAuley, Manling Li, Zecheng He
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 instruction augmentation (refine, remove, swap) to multi-turn data (ablation) | BAGEL-7B process-driven · GenEval | Position: 0.58 -> 0.67 | Improved |
| add instruction augmentation to multi-turn data (ablation) | BAGEL-7B process-driven · GenEval | Color Attributes: 0.50 -> 0.62 | Improved |
| add self-critique after instruction augmentation (ablation) | BAGEL-7B process-driven · GenEval | Position: 0.67 -> 0.72 | Improved |
| add self-critique without augmentation (ablation) | BAGEL-7B process-driven · GenEval | Position: 0.58 -> 0.61 | Improved |
| add scene-graph based critiques (ablation) | BAGEL-7B process-driven · GenEval | Position: 0.67 -> 0.70 | Improved |
| use self-sampled critiques instead of scene-graph critiques (ablation) | BAGEL-7B process-driven · GenEval | Colors / Position / Color Attr.: 0.83 / 0.70 / 0.67 -> 0.87 / 0.72 / 0.69 | Improved |
| add instruction-intermediate conflict supervision (ablation) | BAGEL-7B process-driven · GenEval | Position: 0.66 -> 0.71 | Improved |
| add image-instruction alignment supervision (ablation) | BAGEL-7B process-driven · GenEval | Counting: 0.61 -> 0.73 | Improved |
| add image-instruction alignment supervision (ablation) | BAGEL-7B process-driven · GenEval | Position: 0.66 -> 0.69 | Mixed |
| combine both reasoning subsets (ablation) | BAGEL-7B process-driven · GenEval | Counting / Position: 0.61 -> 0.75; 0.66 -> 0.72 | Improved |
| add process-driven interleaved reasoning fine-tuning to BAGEL-7B | BAGEL-7B · GenEval | Overall: 0.77 -> 0.83 | Improved |
| add process-driven interleaved reasoning fine-tuning to BAGEL | BAGEL · WISE | Overall: 0.70 -> 0.76 | Improved |
Closest papers by meaning
- Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation (CVPR 2026 · similarity 0.91)
- ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.90)
- When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.90)
- GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.90)
- ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.90)
- Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
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.
