SciForma: Structure-Faithful Generation of Scientific Diagrams
SciForma treats diagram generation as a three-part structural test: components, arrows, and text all have to be right at once.
The paper says current open-source models can make plausible scientific diagrams but still fail on errors like reversed arrows or unreadable equations. Its framework adds a structural inventory, a 700K training set, a 2K benchmark, and a post-training method called M-DPO to target the weakest structural axis. The authors report that SciForma-9B beats open-source baselines and GPT-Image-1.5 on SciFormaBench-2K and AIBench. Code and data are promised but not linked in the abstract beyond a placeholder. ArXiv · AI/CL/LG's note
The paper says current open-source models can make plausible scientific diagrams but still fail on errors like reversed arrows or unreadable equations. Its framework adds a structural inventory, a 700K training set, a 2K benchmark, and a post-training method called M-DPO to target the weakest structural axis. The authors report that SciForma-9B beats open-source baselines and GPT-Image-1.5 on SciFormaBench-2K and AIBench. Code and data are promised but not linked in the abstract beyond a placeholder. ArXiv · AI/CL/LG's note
score 5