AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards
AesCode uses image references as aesthetic guidance while keeping code outputs checked against verifiable design properties.
The paper targets visual artifacts like slides, posters, and dashboards, where code is editable but often weak on layout and appearance. AesCode builds a design graph for the canvas, then separates rewards between deterministic checks and a VLM judge using a visual rubric tied to graph elements. The authors report that reference images improve visual quality, and that AesCode-8B and AesCode-32B beat same-scale Qwen3-VL baselines by 31.1 and 22.4 percentage points. HF Daily Papers' note
The paper targets visual artifacts like slides, posters, and dashboards, where code is editable but often weak on layout and appearance. AesCode builds a design graph for the canvas, then separates rewards between deterministic checks and a VLM judge using a visual rubric tied to graph elements. The authors report that reference images improve visual quality, and that AesCode-8B and AesCode-32B beat same-scale Qwen3-VL baselines by 31.1 and 22.4 percentage points. HF Daily Papers' note
score 4