I’m using the NodeJS SDK and am concerned about the efficiency of handling images via base64 encoding. Given that base64 strings are lengthy, will this approach lead to rapid token exhaustion and increased costs compared to more efficient methods? The Python SDK, for example, appears to handle images more optimally with direct PIL object input. Could you clarify if base64 is the most token-efficient method for image input in the NodeJS SDK, and if there are any recommended best practices or alternative approaches to minimize token consumption when working with images?
1 Like
Hi @Preet_Sojitra1,
Welcome to forum, I would say before sending to model you can try counting the number of tokens so that you get a clear understanding how many tokens you are getting consumed even before sending to the model.
For counting multimodal inputs check here: Understand and count tokens | Gemini API | Google AI for Developers
Once you experiment with different images it should you give some confidence, but in general base64 encoding should not increase the number of tokens used.
Documentation use base64 encoded and other way is use upload file api and use that reference inline for prompting. In both cases the number tokens should not defer.
Code is referenced from above count tokens doc
tokens count from base64 encoded string:
tokens count from
Cheers.
1 Like
Verified: Shivam Srivastava