Microsoft's AesCode uses image-based aesthetic rewards to lift code-drawn visuals by 31pt
microsoft · hf · 2026-10-14
AesCode generates information-rich visual artifacts (slides, posters, dashboards) as code, using an image generated from the same prompt as aesthetic guidance. It represents the target as a design graph with attributable properties, deriving decoupled cross-modal rewards — deterministic verifiers for parsable properties, a VLM judge with sample-specific Visual Graph Rubrics for the rest — optimized via RL with per-channel normalization. AesCode-8B and AesCode-32B improve visual quality by 31.1 and 22.4 points over same-scale Qwen3-VL baselines, matching closed-source frontier models like GPT-5.5 and Claude Opus 4.8.
More from Multimodal
- RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding — Guowen Zhang · 2026-10-14
- Gemini 4 Argon builds a live Three.js ice-skating demo running in the browser — alexanderchen · 2026-10-14
- Midjourney style code share: --sref 1618966814 for Van Gogh-style Halloween horror art — tisch_eins · 2026-10-14
- Impasto oil painting prompt shared: thick sculptural brushstroke AI art template — azed_ai · 2026-10-14
- MiniMax H3 keeps face likeness in video clips but loses it in character sheets — diond09 · 2026-10-14
- AI-generated "Stacked Mug" concept: concrete, walnut and copper product design mockup — Delahuntagram · 2026-10-14