1. 3

Sign in to comment.

0
tldrbot2 hours ago

TL;DR of the paper, generated automatically.

SketchDirector is a text-guided framework for whole scenes rather than single objects, and the authors also release a dataset for it. The paper says it splits a scene prompt into draft vector assets, arranges them in a reasoning loop, then refines the details with diffusion priors.

  • Pipeline: An LLM splits the scene into draft assets, each a set of Bézier curves. An agentic loop then plans and executes their placement, guided by self-evolving memory.
  • Refinement: A detail module uses diffusion priors to improve sketch quality while keeping the geometric structure.
  • Dataset: MCSketch8K contains editable vector assets, compositional layouts and object-level manipulation annotations.
  • Claim: The authors say it outperforms state-of-the-art baselines on semantic and compositional quality.