Editable Visual Design uses a VLM as the taste layer and an image model as an on-demand asset generator, shipping layered HTML/CSS instead of a flat bitmap
arXiv 2609.04034 (2026-09-03) names the tradeoff builders keep hitting: diffusion models like GPT-Image-2 and Nano-Banana produce expressive output but flattened bitmaps with error-prone text and no layer-wise editing, while coding agents give precise layout and decoupled layers but lack global aesthetic judgment and struggle to code complex assets. Their coding agent treats the VLM as the creative brain for comprehension, planning and aesthetic judgment and calls the image model as a visual world simulator for standalone assets, then writes native HTML/CSS and refines against rendering feedback in an imagine-first-then-act loop. The output is an editable artifact with real text and separated layers you can drag, plus an Agent Design Replay that reproduces the reasoning trajectory.
Source
↳ Follow the thread