microsoft/AesCode-8B
AesCode-8B by Microsoft is an 8 billion parameter multimodal language model, based on Qwen3-VL-8B-Instruct, designed for generating information-rich visual artifacts as structured, editable, and verifiable HTML/CSS. It excels at creating slides, posters, and dashboards by using an image as an aesthetic reference while adhering to content prompts, overcoming challenges faced by traditional code models and image generators. The model achieves superior performance in generating visually compelling and semantically accurate web content, particularly when a reference image is provided, and is trained with graph-structured supervision and decoupled cross-modal rewards.
Loading preview...
AesCode-8B Overview
AesCode-8B, developed by Microsoft, is an 8 billion parameter multimodal model built upon the Qwen3-VL-8B-Instruct architecture. Its primary function is to generate information-rich visual artifacts, such as slides, posters, and dashboards, as structured, editable, and verifiable HTML/CSS. Unlike traditional code models that struggle with visual layout or image generators that misrender text, AesCode-8B leverages an image-generated reference for aesthetic guidance while strictly following content prompts.
Key Capabilities
- Structured HTML/CSS Generation: Produces editable and verifiable web content for visual artifacts.
- Multimodal Input: Utilizes both text prompts and optional reference images to guide generation, separating semantic requirements from visual cues.
- Enhanced Visual and Rule Adherence: Achieves high scores in both programmatic checks (Text, Bound, Chart) and visual quality (Content, Layout, Style).
- Advanced Training Methodology: Employs cold-start Supervised Fine-Tuning (SFT) followed by GDPO (Generalized Distributional Policy Optimization) across seven distinct reward channels, including execution, text, boundary, tablechart, layout, whitespace, and design.
- Robust Performance: Outperforms reference-conditioned GPT-5.5 and Claude Opus 4.8 in overall scores for infographic generation, demonstrating significant improvements over its Qwen3-VL-8B backbone.
Good For
- Generating editable visual documents: Ideal for creating slides, posters, dashboards, and reports that require structured and verifiable output.
- Research in multimodal code generation: Provides a strong foundation for exploring how to combine visual and semantic understanding in code generation.
- Verifiable visual design: Ensures that generated designs are both aesthetically pleasing and programmatically correct.
- Applications requiring visual planning: Particularly effective when a reference image can be supplied to guide layout and style, though it maintains strong performance even without one.