SolveEdit: Benchmarking Visual Problem Solving in Generative Models
The benchmark tests whether image generators can infer and carry out goal-driven scene edits without breaking what should stay unchanged.
SolveEdit contains 2,728 visual problem-solving cases built around scene transformation. Models must understand the image and goal, decide what change is valid, and preserve unrelated content. The paper says the strongest evaluated model reaches 57.0% SolveScore. A two-stage planner, SolveEdit-PLAN, improves scores across three generators, including GPT-Image-2 rising from 57.0% to 71.6% under the matched evaluation. HF Daily Papers' note
SolveEdit contains 2,728 visual problem-solving cases built around scene transformation. Models must understand the image and goal, decide what change is valid, and preserve unrelated content. The paper says the strongest evaluated model reaches 57.0% SolveScore. A two-stage planner, SolveEdit-PLAN, improves scores across three generators, including GPT-Image-2 rising from 57.0% to 71.6% under the matched evaluation. HF Daily Papers' note
score 4