Research map / Image generation / Text-to-image diffusion
Scientific Image Synthesis: Benchmarking, Methodologies, and Downstream Utility
honglin lin, Chonghan Qin, Zheng Liu, Qizhi Pei, Yu Li, Zhanping Zhong, Xin Gao, Wei Li, Wentao Zhang, Yanfeng Wang, Conghui He, Lijun Wu
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 |
|---|---|---|---|
| generate science images via code (ImgCoder) instead of pixel T2I with the same Qwen family | Qwen3-ImgCoder vs Qwen-Image · SciGenBench | Rinv (%): 38.86 -> 56.38 | Improved |
| use Gemini-3-Pro-ImgCoder instead of best pixel model Nanobanana-Pro | scientific image generation · SciGenBench | Rinv (%): 73.41 -> 77.87 | Improved |
| code-driven vs pixel-based generation on biology | Qwen3-ImgCoder · SciGenBench Biology | Rinv (%): Qwen-Image 43.14 vs Qwen3-ImgCoder 41.18 | Hurt |
| code-driven vs pixel-based generation on image fidelity metrics | ImgCoder · SciGenBench-SeePhys | FID: GPT-Image-1 77.31 vs Gemini-3-Pro-ImgCoder 107.67 | Hurt |
| RL fine-tune LMM on verified synthetic scientific images | Qwen3-VL-8B-Instruct · GEO3K and MathVision average | average accuracy: 54.5 -> 58.2 | Improved |
| filter incorrect synthetic images from training data (ablation) | Qwen-Image data · GEO3K+MathVision average | average accuracy: 57.1 -> 57.8 | Improved |
| filter incorrect synthetic images from training data (ablation) | Nanobanana-Pro data · GEO3K / MathVision | accuracy: GEO3K 70.1 -> 68.7; MV 46.1 -> 47.7; AVG 58.1 -> 58.2 | Mixed |
| use higher-quality generator (Nanobanana-Pro instead of Qwen-Image) for training images (ablation) | RL data · GEO3K+MathVision average | average accuracy: 57.1 -> 58.1 | Improved |
| scale synthetic training set from 50 to 1.4K samples (ablation) | Nanobanana-Pro images · downstream accuracy | accuracy: 43.9 -> 46.1 | Improved |
| use code-generated (ImgCoder) vs pixel images for training (ablation) | Qwen3-VL-8B RL · MathVision | accuracy: Qwen-Image 45.9 -> Qwen-Imgcoder 46.9 | Improved |
| use ImgCoder vs pixel on GEO3K training data (ablation) | Qwen3-VL-8B RL · GEO3K | accuracy: Qwen-Image 68.2 vs Qwen-Imgcoder 67.9 | No clear effect |
Closest papers by meaning
- MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning (NeurIPS 2025 · similarity 0.90)
- DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.89)
- ScImage: How good are multimodal large language models at scientific text-to-image generation? (ICLR 2025 · similarity 0.89)
- ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.89)
- GGBench: A Geometric Generative Reasoning Benchmark for Unified Multimodal Models (CVPR 2026 · similarity 0.89)
- Generative Universal Verifier as Multimodal Meta-Reasoner (ICLR 2026 · 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.
