SciForma: Structure-Faithful Generation of Scientific Diagrams
SciForma trains diagram generation against component, arrow, and text correctness at the same time.
The paper argues that scientific methodology diagrams fail if any one structural element is wrong, such as a reversed arrow or unreadable equation. SciForma uses a structural inventory, a 700K-example training set, and a 2K-example benchmark to target those failure modes separately. Its M-DPO method routes post-training pressure toward the weakest structural dimension instead of collapsing quality into one reward score. The authors say SciForma-9B beats open-source baselines and GPT-Image-1.5 on their diagram benchmarks. HF Daily Papers' note
The paper argues that scientific methodology diagrams fail if any one structural element is wrong, such as a reversed arrow or unreadable equation. SciForma uses a structural inventory, a 700K-example training set, and a 2K-example benchmark to target those failure modes separately. Its M-DPO method routes post-training pressure toward the weakest structural dimension instead of collapsing quality into one reward score. The authors say SciForma-9B beats open-source baselines and GPT-Image-1.5 on their diagram benchmarks. HF Daily Papers' note
score 5