Step Image Edit 2 tops KRIS-Bench at $0.003 per image

StepFun released Step Image Edit 2, a 3.5-billion-parameter image model from the StepFun lab, claiming the top score on KRIS-Bench, a published benchmark that tests whether an image edit correctly applies factual, conceptual and procedural knowledge rather than just looking plausible. The benchmark's research paper scores edits across 22 task types built from those three knowledge categories; StepFun says Step Image Edit 2 ranks first specifically in the overall, factual and conceptual categories, meaning it has to get real-world facts and concept-based edits right in those areas, not only match visual quality.

That top ranking is StepFun's own report of its score against the public benchmark, not an outcome confirmed on an independently maintained leaderboard, so it should be read as a claim rather than a verified standing. The figures StepFun disclosed directly are more concrete: a 3.5B-parameter model generating an image from text in 0.7 seconds, applying an edit in 1.6 seconds, and pricing each image at $0.003, alongside bilingual Chinese and English text rendering that the company says survives an edit.