vlm-judge

AesCode introduces a framework that generates visual artifacts (slides, posters, dashboards) as code by combining content-grounded code generation with aesthetic guidance from reference images, using decoupled cross-modal rewards optimized via RL. AesCode-8B and AesCode-32B improve visual quality by 31.1 and 22.4 percentage points over Qwen3-VL baselines, reaching parity with frontier closed-source models.