I reduced image-processing token usage by ~95% compared with GPT-4o direct vision, while maintaining roughly the same accuracy.How significant is that?[P]

I'm testing a new approach for reducing the cost of image-based LLM inference. I evaluated it on the MOMA Graph benchmark, using 1,315 questions. Compared with using GPT-4o to process the original images directly, I observed approximately: ~95% lower token usage roughly the same accuracy as the GPT-