Much like UML and BPMN give software and process diagrams a compact vocabulary of typed nodes and edges, SciForma formalizes scientific figures as a structural inventory over three independently verifiable primitives, Component, Arrow, and Text (C/A/T). The same inventory drives training captions and benchmark rubrics, so what the model is asked to draw is exactly what the judge scores.

Correctness is conjunctive, meaning a figure counts as correct only when all three axes are simultaneously right. A missing arrow or mislabelled block is not offset by a clean layout, so the model cannot coast on averaged scores and must render every axis faithfully.

726K methodology diagrams from 593K arXiv LaTeX sources (2015–2025), kept after consensus filtering by Qwen3-VL and OpenAI-o3. Each figure carries a structured dense description covering global layout, module attributes, and topological dependencies. 71K editing triplets are mined by SAM3 element segmentation with OpenAI-o3 written grounded instructions.

Data pipeline: extraction → detection → caption generation → filtering → editing augmentation.

2,000 held-out prompts stratified by structural complexity and deliberately weighted toward Medium and Hard, where dense multi-branch diagrams remain the real challenge. Each sample carries an element checklist extracted by GPT-5.4 and corrected by human annotators, then scored on the C/A/T axes with severity-weighted error counting.

Per-axis rubric with critical and moderate error annotations.

SciForma fine-tunes FLUX.2-klein-base-9B on SciFormaPile in two stages. Stage 1 trains all 9B parameters on the full 656K generation pairs with resolution bucketing to keep batches spatially homogeneous across widely varying aspect ratios. Stage 2 initializes from Stage-1 EMA weights and co-trains on 244K High-quality pairs together with the 71K editing triplets, so a single checkpoint serves both generation and editing through a concat-then-prune design that adds no new parameters.

SFT leaves axis-specific artifacts such as missing components, garbled labels, and broken arrows, and these failure modes are near-independent across the three axes (inter-axis ρ < 0.2). MD³PO pairs each winner with dimension-anchored losers, one per axis, and couples them through an InfoNCE-style contrastive objective. The loss only vanishes when the policy prefers the winner on every axis simultaneously, and a softmax over per-axis reward gaps automatically concentrates gradient on whichever dimension is currently weakest.

M-DPO ablation on SciFormaBench-2K. Text axis benefits most (+3.16 pts).

At inference time a VLM critic inspects the generated diagram against its structural inventory and flags concrete defects, which are compiled into a priority-ranked edit plan. SciForma-9B applies each atomic edit through the same editing branch trained in Stage 2, and the loop repeats for a fixed number of rounds or stops early once the critic reports no remaining structural errors, yielding the closed-loop refinement shown below.

Closed-loop editor. Defect detection, priority-ranked edit plan, SciForma-9B targeted inpainting, global verification. The loop terminates when the structural inventory is satisfied.