SmolVLM WebGPU

I tried the demo on https://huggingface.co/spaces/HuggingFaceTB/SmolVLM-256M-Instruct-WebGPU and also ran it locally, but the outputs were different. For example，this image ("https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/new-york.jpg"), which only displays "here is the statue of liberty" locally, can you give me some hints？