Ueberblick
October 2026 paper using generated reference images to guide editable HTML/CSS generation for dashboards, webpages, slides, and posters. AesCode keeps the prompt authoritative for content, represents the design as a graph, and trains with separate programmatic checks and visual rewards.
Zusammenfassung
Why AesCode Adds New Signal:
1. Guidance without copying errors: Learns layout and style from generated images while grounding text and data in the prompt
2. Attributable rewards: Routes design-graph properties to deterministic verifiers or graph-specific visual questions, then normalizes reward channels separately
3. Separated evaluation: Reports visual quality and verifiable correctness on 300 infographic prompts with three generations each and no selection
4. Research boundary: Evaluates visual artifacts rather than runtime interaction; complex relationships and SVG icons remain difficult, and reference images can contain hallucinations
[Official project, examples, code, and model links](https://microsoft.github.io/AesCode/).
1. Guidance without copying errors: Learns layout and style from generated images while grounding text and data in the prompt
2. Attributable rewards: Routes design-graph properties to deterministic verifiers or graph-specific visual questions, then normalizes reward channels separately
3. Separated evaluation: Reports visual quality and verifiable correctness on 300 infographic prompts with three generations each and no selection
4. Research boundary: Evaluates visual artifacts rather than runtime interaction; complex relationships and SVG icons remain difficult, and reference images can contain hallucinations
[Official project, examples, code, and model links](https://microsoft.github.io/AesCode/).
Tags
ui-generationvisual-guidancedesign-graphsreinforcement-learningcode-verification