ScribbleEdit: A Benchmark for Scribble-Only Image Editing

arXiv cs.GR details ScribbleEdit, a benchmark built to measure how well image-editing models can follow scribble-only instructions. The goal is practical: interactive tools often rely on quick sketch-like input, but today’s VLM- and LLM-based editors still have trouble turning those sparse marks into the intended edit.

The paper introduces an automated data-construction pipeline and a dedicated evaluation protocol that focuses on intention alignment, not just pixel-level similarity. That matters for developers building creative tools, because a model can produce a visually plausible edit while still missing the user’s actual intent.

The authors say existing editing systems fail to reliably capture scribble semantics, which suggests the bottleneck is not just spatial reasoning but also understanding what the scribble is meant to represent. To push the task forward, they also propose a soft-token baseline that improves semantic understanding of scribbles and outperforms standard image-editing models on ScribbleEdit.

For teams working on generative editing, this is a useful reminder that lightweight input modes need their own benchmarks. If scribble-driven workflows are part of your roadmap, the evaluation setup here could be a good reference point for measuring whether a model is actually usable in production.