VISA is an agentic framework that treats multimodal instruction data synthesis as an iterative, closed-loop process. At each round, it analyzes images to filter incompatible constraints and discover new ones, samples diversity- and difficulty-aware constraint sets from persistent memory, generates candidate instructions, and verifies them using executable tools and structured LLM judges.
Failed samples trigger diagnostic-guided recovery, while accepted samples are probed against the target model to estimate difficulty. The outcomes are written to memory, which adapts the next round's focus on unresolved weaknesses. Experiments on the MM-IFEval benchmark show consistent improvements in instruction following over strong baselines, while general multimodal capability is preserved across seven public benchmarks.