A new arXiv preprint, "Pixels or Code? Benchmarking Instruction-Guided Visual Artifact Operation," tackles the challenge of editing complex visual artifacts such as charts, technical diagrams, webpages, and document pages. The abstract notes that these artifacts underpin many professional workflows, yet editing them reliably requires more than simple image manipulation—it demands direct modification and updates that respect dependencies among elements.
The paper's title signals its central comparison: whether AI systems should edit such artifacts at the pixel level or by generating and modifying code. The abstract itself is brief, but it frames the benchmark as instruction-guided, meaning models must follow user commands to perform operations on the artifacts.
Because only the abstract is available, details about the benchmark's construction, evaluation metrics, or results are not yet public. The contribution, as described, is a benchmark designed to measure how well different approaches handle these dependency-aware edits—an open question for anyone building tools for document, diagram, or webpage manipulation.