RULER: Instance-aware Rubric Rewards for SVG Generation
RULER uses per-instruction rubric rewards so reinforcement learning improves text-to-SVG generation without ground truth.
RULER is a reinforcement-learning method that scores SVG generation with instance-aware rubrics instead of scalar image metrics that transfer poorly to vector art. Each instruction becomes a six-item rubric covering semantic, visual, and stylistic axes; a vision-language judge scores rendered rollouts, and weighted scores train the policy with Group Relative Policy Optimization. No paired SVG ground truth or human preference labels are required. On MMSVG-Illustration and MMSVG-Icon, rubric scores rise from 0.432 and 0.395 to 0.693 and 0.683, matching much larger DeepSeek-V3.
- RULER builds a six-item instance-aware rubric from each text instruction alone.
- A vision-language judge scores rendered SVG rollouts item by item for Group Relative Policy Optimization.
- Rubric scores rise from 0.432/0.395 to 0.693/0.683 on MMSVG-Illustration and MMSVG-Icon.
- Results beat dedicated SVG specialists and match the much larger DeepSeek-V3.
Full article201 words · extracted from huggingface.co · click to collapse
Generating Scalable Vector Graphics (SVG) code from natural-language instructions is an open-ended task without absolute visual ground truth, leaving both evaluation and policy optimization without a faithful signal. Scalar metrics (CLIP, Aesthetic) calibrated on natural images transfer poorly to stylized vector content, and reusing them as RL rewards triggers reward hacking. We address both limitations with rubric-based scoring. We first establish empirically that prompting a vision-language judge with a multi-axis rubric correlates with human judgments far better than scalar metrics, both across samples and within instructions. Building on this finding, we introduce RULER (Instance-aware Rubric Rewards for Reinforcement Learning), which converts each instruction into an instance-aware rubric of six items spanning semantic, visual, and stylistic axes; a judge VLM scores rendered rollouts item-by-item, and the weighted satisfactions form a fine-grained reward optimized via Group Relative Policy Optimization. Because the rubric is derived from text alone, RULER requires neither paired SVG ground truth nor human preference labels. On MMSVG-Illustration and MMSVG-Icon, RULER lifts the rubric score from 0.432/0.395 to 0.693/0.683, surpassing dedicated SVG specialists and matching the substantially larger DeepSeek-V3, with ablations identifying rubric design as the active lever for RL on open-ended SVG generation. The project page is available at https://hangyuran.github.io/RULER/.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.25270