Making AI vision models run faster by cutting useless image chunks — no retraining required

Researchers found a way to speed up large vision-language models by automatically removing redundant visual information before the model processes it, without having to retrain the model from scratch. In practice: the same AI output, but 30-40% faster and cheaper to run.

Vision-language models are getting slower as they get smarter — they now process thousands of image tokens per query, which costs real money in compute. This method identifies which tokens don't add new information and discards them before processing even starts. That means companies can deploy these models at lower cost without rebuilding them, which matters because the cost curve is what determines whether a technology gets used at scale or stays expensive enough to lock out smaller competitors.