Attractive editing is not necessarily correct editing
The ABACUS paper, first posted on October 20, 2026, proposes checking exact object counts and conservation of the original scene together. An instruction to add three objects should not pass if an editor instead repurposes existing objects or removes something elsewhere. Its arXiv record states acceptance at WACV 2027 in the Evaluations & Datasets track. This is an evaluation development, not the launch of another image generator.
Test the counter as well as the editor
The protocol uses instance-level segmentation ground truth and detector agreement. A VLM does not count objects or assign the edit score, although it can assist with category confirmation for rare ambiguous crops. Calling the entire pipeline language-model-free would therefore be inaccurate.
Its scope is the 80 COCO object categories, not every possible editing task. Crucially, the hosted-API experiments were run in June and August 2026; the October paper does not establish today's ranking of the latest models. For production, separate visual appeal, requested count, retained objects and background changes in the review checklist. When quantities matter in product or instructional imagery, a final human count remains appropriate even after an automated check.